The present and future of AI

Finale doshi-velez on how ai is shaping our lives and how we can shape ai.

image of Finale Doshi-Velez, the John L. Loeb Professor of Engineering and Applied Sciences

Finale Doshi-Velez, the John L. Loeb Professor of Engineering and Applied Sciences. (Photo courtesy of Eliza Grinnell/Harvard SEAS)

How has artificial intelligence changed and shaped our world over the last five years? How will AI continue to impact our lives in the coming years? Those were the questions addressed in the most recent report from the One Hundred Year Study on Artificial Intelligence (AI100), an ongoing project hosted at Stanford University, that will study the status of AI technology and its impacts on the world over the next 100 years.

The 2021 report is the second in a series that will be released every five years until 2116. Titled “Gathering Strength, Gathering Storms,” the report explores the various ways AI is  increasingly touching people’s lives in settings that range from  movie recommendations  and  voice assistants  to  autonomous driving  and  automated medical diagnoses .

Barbara Grosz , the Higgins Research Professor of Natural Sciences at the Harvard John A. Paulson School of Engineering and Applied Sciences (SEAS) is a member of the standing committee overseeing the AI100 project and Finale Doshi-Velez , Gordon McKay Professor of Computer Science, is part of the panel of interdisciplinary researchers who wrote this year’s report. 

We spoke with Doshi-Velez about the report, what it says about the role AI is currently playing in our lives, and how it will change in the future.  

Q: Let's start with a snapshot: What is the current state of AI and its potential?

Doshi-Velez: Some of the biggest changes in the last five years have been how well AIs now perform in large data regimes on specific types of tasks.  We've seen [DeepMind’s] AlphaZero become the best Go player entirely through self-play, and everyday uses of AI such as grammar checks and autocomplete, automatic personal photo organization and search, and speech recognition become commonplace for large numbers of people.  

In terms of potential, I'm most excited about AIs that might augment and assist people.  They can be used to drive insights in drug discovery, help with decision making such as identifying a menu of likely treatment options for patients, and provide basic assistance, such as lane keeping while driving or text-to-speech based on images from a phone for the visually impaired.  In many situations, people and AIs have complementary strengths. I think we're getting closer to unlocking the potential of people and AI teams.

There's a much greater recognition that we should not be waiting for AI tools to become mainstream before making sure they are ethical.

Q: Over the course of 100 years, these reports will tell the story of AI and its evolving role in society. Even though there have only been two reports, what's the story so far?

There's actually a lot of change even in five years.  The first report is fairly rosy.  For example, it mentions how algorithmic risk assessments may mitigate the human biases of judges.  The second has a much more mixed view.  I think this comes from the fact that as AI tools have come into the mainstream — both in higher stakes and everyday settings — we are appropriately much less willing to tolerate flaws, especially discriminatory ones. There's also been questions of information and disinformation control as people get their news, social media, and entertainment via searches and rankings personalized to them. So, there's a much greater recognition that we should not be waiting for AI tools to become mainstream before making sure they are ethical.

Q: What is the responsibility of institutes of higher education in preparing students and the next generation of computer scientists for the future of AI and its impact on society?

First, I'll say that the need to understand the basics of AI and data science starts much earlier than higher education!  Children are being exposed to AIs as soon as they click on videos on YouTube or browse photo albums. They need to understand aspects of AI such as how their actions affect future recommendations.

But for computer science students in college, I think a key thing that future engineers need to realize is when to demand input and how to talk across disciplinary boundaries to get at often difficult-to-quantify notions of safety, equity, fairness, etc.  I'm really excited that Harvard has the Embedded EthiCS program to provide some of this education.  Of course, this is an addition to standard good engineering practices like building robust models, validating them, and so forth, which is all a bit harder with AI.

I think a key thing that future engineers need to realize is when to demand input and how to talk across disciplinary boundaries to get at often difficult-to-quantify notions of safety, equity, fairness, etc. 

Q: Your work focuses on machine learning with applications to healthcare, which is also an area of focus of this report. What is the state of AI in healthcare? 

A lot of AI in healthcare has been on the business end, used for optimizing billing, scheduling surgeries, that sort of thing.  When it comes to AI for better patient care, which is what we usually think about, there are few legal, regulatory, and financial incentives to do so, and many disincentives. Still, there's been slow but steady integration of AI-based tools, often in the form of risk scoring and alert systems.

In the near future, two applications that I'm really excited about are triage in low-resource settings — having AIs do initial reads of pathology slides, for example, if there are not enough pathologists, or get an initial check of whether a mole looks suspicious — and ways in which AIs can help identify promising treatment options for discussion with a clinician team and patient.

Q: Any predictions for the next report?

I'll be keen to see where currently nascent AI regulation initiatives have gotten to. Accountability is such a difficult question in AI,  it's tricky to nurture both innovation and basic protections.  Perhaps the most important innovation will be in approaches for AI accountability.

Topics: AI / Machine Learning , Computer Science

Cutting-edge science delivered direct to your inbox.

Join the Harvard SEAS mailing list.

Scientist Profiles

Finale Doshi-Velez

Finale Doshi-Velez

Herchel Smith Professor of Computer Science

Press Contact

Leah Burrows | 617-496-1351 | [email protected]

Related News

Four people standing in a line, one wearing a Harvard sweatshirt, two holding second-place statues

A new phase for Harvard Quantum Computing Club

SEAS students place second at MIT quantum hackathon

Computer Science , Quantum Engineering , Undergraduate Student Profile

H.T. Kung

Kung elected to ACM

Computer scientist honored for transformative contributions

Awards , Computer Science

image of Ariel Procaccia

Procaccia elected to the Association for the Advancement of Artificial Intelligence

Computer Scientist honored for contributions to AI

Advertisement

Advertisement

Artificial Intelligence in Education (AIEd): a high-level academic and industry note 2021

  • Original Research
  • Open access
  • Published: 07 July 2021
  • Volume 2 , pages 157–165, ( 2022 )

Cite this article

You have full access to this open access article

  • Muhammad Ali Chaudhry   ORCID: orcid.org/0000-0003-0154-2613 1 &
  • Emre Kazim 2  

33k Accesses

43 Citations

24 Altmetric

Explore all metrics

In the past few decades, technology has completely transformed the world around us. Indeed, experts believe that the next big digital transformation in how we live, communicate, work, trade and learn will be driven by Artificial Intelligence (AI) [ 83 ]. This paper presents a high-level industrial and academic overview of AI in Education (AIEd). It presents the focus of latest research in AIEd on reducing teachers’ workload, contextualized learning for students, revolutionizing assessments and developments in intelligent tutoring systems. It also discusses the ethical dimension of AIEd and the potential impact of the Covid-19 pandemic on the future of AIEd’s research and practice. The intended readership of this article is policy makers and institutional leaders who are looking for an introductory state of play in AIEd.

Similar content being viewed by others

research paper based on artificial intelligence

Artificial intelligence in education: Addressing ethical challenges in K-12 settings

Selin Akgun & Christine Greenhow

research paper based on artificial intelligence

Evolution and Revolution in Artificial Intelligence in Education

Ido Roll & Ruth Wylie

Ethical principles for artificial intelligence in education

Andy Nguyen, Ha Ngan Ngo, … Bich-Phuong Thi Nguyen

Avoid common mistakes on your manuscript.

1 Introduction

Artificial Intelligence (AI) is changing the world around us [ 42 ]. As a term it is difficult to define even for experts because of its interdisciplinary nature and evolving capabilities. In the context of this paper, we define AI as a computer system that can achieve a particular task through certain capabilities (like speech or vision) and intelligent behaviour that was once considered unique to humans [ 54 ]. In more lay terms we use the term AI to refer to intelligent systems that can automate tasks traditionally carried out by humans. Indeed, we read AI within the continuation of the digital age, with increased digital transformation changing the ways in which we live in the world. With such change the skills and knowhow of people must reflect the new reality and within this context, the World Economic Forum identified sixteen skills, referred to as twenty-first century skills necessary for the future workforce [ 79 ]. This includes skills such as technology literacy, communication, leadership, curiosity, adaptability, etc. These skills have always been important for a successful career, however, with the accelerated digital transformation of the past 2 years and the focus on continuous learning in most professional careers, these skills are becoming necessary for learners.

AI will play a very important role in how we teach and learn these new skills. In one dimension, ‘AIEd’ has the potential to dramatically automate and help track the learner’s progress in all these skills and identify where best a human teacher’s assistance is needed. For teachers, AIEd can potentially be used to help identify the most effective teaching methods based on students’ contexts and learning background. It can automate monotonous operational tasks, generate assessments and automate grading and feedback. AI does not only impact what students learn through recommendations, but also how they learn, what are the learning gaps, which pedagogies are more effective and how to retain learner’s attention. In these cases, teachers are the ‘human-in-the-loop’, where in such contexts, the role of AI is only to enable more informed decision making by teachers, by providing them predictions about students' performance or recommending relevant content to students after teachers' approval. Here, the final decision makers are teachers.

Segal et al. [ 58 ] developed a system named SAGLET that utilized ‘human-in-the-loop’ approach to visualize and model students’ activities to teachers in real-time enabling them to intervene more effectively as and when needed. Here the role of AI is to empower the teachers enabling them to enhance students’ learning outcomes. Similarly, Rodriguez et al. [ 52 ] have shown how teachers as ‘human-in-the-loop’ can customize multimodal learning analytics and make them more effective in blended learning environments.

Critically, all these achievements are completely dependent on the quality of available learner data which has been a long-lasting challenge for ed-tech companies, at least until the pandemic. Use of technology in educational institutions around the globe is increasing [ 77 ], however, educational technology (ed-tech) companies building AI powered products have always complained about the lack of relevant data for training algorithms. The advent and spread of Covid in 2019 around the world pushed educational institutions online and left them at the mercy of ed-tech products to organize content, manage operations, and communicate with students. This shift has started generating huge amounts of data for ed-tech companies on which they can build AI systems. According to a joint report: ‘Shock to the System’, published by Educate Ventures and Cambridge University, optimism of ed-tech companies about their own future increased during the pandemic and their most pressing concern became recruitment of too many customers to serve effectively [ 15 ].

Additionally, most of the products and solutions provided by ed-tech start-ups lack the quality and resilience to cope with intensive use of several thousands of users. Product maturity is not ready for huge and intense demand as discussed in Sect. “ Latest research ” below. We also discuss some of these products in detail in Sect. “ Industry’s focus ” below. How do we mitigate the risks of these AI powered products and who monitors the risk? (we return to this theme in our discussion of ethics—Sect. “ Ethical AIEd ”).

This paper is a non-exhaustive overview of AI in Education that presents a brief survey of the latest developments of AI in Education. It begins by discussing different aspects of education and learning where AI is being utilized, then turns to where we see the industry’s current focus and then closes with a note on ethical concerns regarding AI in Education. This paper also briefly evaluates the potential impact of the pandemic on AI’s application in education. The intended readership of this article is the policy community and institutional executives seeking an instructive introduction to the state of play in AIEd. The paper can also be read as a rapid introduction to the state of play of the field.

2 Latest research

Most work within AIEd can be divided into four main subdomains. In this section, we survey some of the latest work in each of these domains as case studies:

Reducing teachers’ workload: the purpose of AI in Education is to reduce teachers’ workload without impacting learning outcomes

Contextualized learning for students: as every learner has unique learning needs, the purpose of AI in Education is to provide customized and/or personalised learning experiences to students based on their contexts and learning backgrounds.

Revolutionizing assessments: the purpose of AI in Education is to enhance our understanding of learners. This not only includes what they know, but also how they learn and which pedagogies work for them.

Intelligent tutoring systems (ITS): the purpose of AI in Education is to provide intelligent learning environments that can interact with students, provide customized feedback and enhance their understanding of certain topics

2.1 Reducing teachers’ workload

Recent research in AIEd is focusing more on teachers than other stakeholders of educational institutions, and this is for the right reasons. Teachers are at the epicenter of every learning environment, face to face or virtual. Participatory design methodologies ensure that teachers are an integral part of the design of new AIEd tools, along with parents and learners [ 45 ]. Reducing teachers’ workload has been a long-lasting challenge for educationists, hoping to achieve more affective teaching in classrooms by empowering the teachers and having them focus more on teaching than the surrounding activities.

With the focus on online education during the pandemic and emergence of new tools to facilitate online learning, there is a growing need for teachers to adapt to these changes. Importantly, teachers themselves are having to re-skill and up-skill to adapt to this age, i.e. the new skills that teachers need to develop to fully utilize the benefits of AIEd [ 39 ]. First, they need to become tech savvy to understand, evaluate and adapt new ed-tech tools as they become available. They may not necessarily use these tools, but it is important to have an understanding of what these tools offer and if they share teachers’ workload. For example, Zoom video calling has been widely used during the pandemic to deliver lessons remotely. Teachers need to know not only how to schedule lessons on Zoom, but also how to utilize functionalities like breakout rooms to conduct group work and Whiteboard for free style writing. Second, teachers will also need to develop analytical skills to interpret the data that are visualized by these ed-tech tools and to identify what kind of data and analytics tools they need to develop a better understanding of learners. This will enable teachers to get what they exactly need from ed-tech companies and ease their workload. Third, teachers will also need to develop new team working, group and management skills to accommodate new tools in their daily routines. They will be responsible for managing these new resources most efficiently.

Selwood and Pilkington [ 61 ] showed that the use of Information and Communication Technologies (ICT) leads to a reduction in teachers’ workload if they use it frequently, receive proper training on how to use ICT and have access to ICT in home and school. During the pandemic, teachers have been left with no options other than online teaching. Spoel et al. [ 76 ] have shown that the previous experience with ICT did not play a significant role in how they dealt with the online transition during pandemic. Suggesting that the new technologies are not a burden for teachers. It is early to draw any conclusions on the long-term effects of the pandemic on education, online learning and teachers’ workload. Use of ICT during the pandemic may not necessarily reduce teacher workload, but change its dynamics.

2.2 Contextualized learning for students

Every learner has unique learning contexts based on their prior knowledge about the topic, social background, economic well-being and emotional state [ 41 ]. Teaching is most effective when tailored to these changing contexts. AIEd can help in identifying the learning gaps in each learner, offer content recommendations based on that and provide step by step solutions to complex problems. For example, iTalk2Learn is an opensource platform that was developed by researchers to support math learning among students between 5 and 11 years of age [ 22 ]. This tutor interacted with students through speech, identified when students were struggling with fractions and intervened accordingly. Similarly, Pearson has launched a calculus learning tool called AIDA that provides step by step guidance to students and helps them complete calculus tasks. Use of such tools by young students also raises interesting questions about the illusion of empathy that learners may develop towards such educational bots [ 73 ].

Open Learner Models [ 12 , 18 ] have been widely used in AIEd to facilitate learners, teachers and parents in understanding what learners know, how they learn and how AI is being used to enhance learning. Another important construct in understanding learners is self-regulated learning [ 10 , 68 ]. Zimmerman and Schunk [ 85 ] define self-regulated learning as learner’s thoughts, feelings and actions towards achieving a certain goal. Better understanding of learners through open learner models and self-regulated learning is the first step towards contextualized learning in AIEd. Currently, we do not have completely autonomous digital tutors like Amazon’s Alexa or Apple’s Siri for education but domain specific Intelligent Tutoring Systems (ITS) are also very helpful in identifying how much students know, where they need help and what type of pedagogies would work for them.

There are a number of ed-tech tools available to develop basic literacy skills in learners like double digit division or improving English grammar. In future, AIEd powered tools will move beyond basic literacy to develop twenty-first century skills like curiosity [ 49 ], initiative and creativity [ 51 ], collaboration and adaptability [ 36 ].

2.3 Revolutionizing assessments

Assessment in educational context refers to ‘any appraisal (or judgement or evaluation) of a student’s work or performance’ [ 56 ]. Hill and Barber [ 27 ] have identified assessments as one of the three pillars of schooling along with curriculum and learning and teaching. The purpose of modern assessments is to evaluate what students know, understand and can do. Ideally, assessments should take account of the full range of student abilities and provide useful information about learning outcomes. However, every learner is unique and so are their learning paths. How can standardized assessment be used to evaluate every student, with distinct capabilities, passions and expertise is a question that can be posed to broader notions of educational assessment. According to Luckin [ 37 ] from University College London, ‘AI would provide a fairer, richer assessment system that would evaluate students across a longer period of time and from an evidence-based, value-added perspective’.

AIAssess is an example of an intelligent assessment tool that was developed by researchers at UCL Knowledge lab [ 38 , 43 ]. It assessed students learning math and science based on three models: knowledge model, analytics model and student model. Knowledge component stored the knowledge about each topic, the analytics component analyzed students’ interactions and the student model tracked students’ progress on a particular topic. Similarly, Samarakou et al. [ 57 ] have developed an AI assessment tool that also does qualitative evaluation of students to reduce the workload of instructors who would otherwise spend hours evaluating every exercise. Such tools can be further empowered by machine learning techniques such as semantic analysis, voice recognition, natural language processing and reinforcement learning to improve the quality of assessments.

2.4 Intelligent tutoring systems (ITS)

An intelligent tutoring system is a computer program that tries to mimic a human teacher to provide personalized learning to students [ 46 , 55 ]. The concept of ITS in AIEd is decades old [ 9 ]. There have always been huge expectations from ITS capabilities to support learning. Over the years, we have observed that there has been a significant contrast between what ITS were envisioned to deliver and what they have actually been capable of doing [ 4 ].

A unique combination of domain models [ 78 ], pedagogical models [ 44 ] and learner models [ 20 ] were expected to provide contextualized learning experiences to students with customized content, like expert human teachers [ 26 , 59 , 65 ],. Later, more models were introduced to enhance students' learning experience like strategy model, knowledge-base model and communication model [ 7 ]. It was expected that an intelligent tutoring system would not just teach, but also ensure that students have learned. It would care for students [ 17 ]. Similar to human teachers, ITS would improve with time. They would learn from their experiences, ‘understand’ what works in which contexts and then help students accordingly [ 8 , 60 ].

In recent years, ITS have mostly been subject and topic specific like ASSISTments [ 25 ], iTalk2Learn [ 23 ] and Aida Calculus. Despite being limited in terms of the domain that a particular intelligent tutoring system addresses, they have proven to be effective in providing relevant content to students, interacting with students [ 6 ] and improving students’ academic performance [ 18 , 41 ]. It is not necessary that ITS would work in every context and facilitate every teacher [ 7 , 13 , 46 , 48 ]. Utterberg et al. [78] showed why teachers have abandoned technology in some instances because it was counterproductive. They conducted a formative intervention with sixteen secondary school mathematics teachers and found systemic contradictions between teachers’ opinions and ITS recommendations, eventually leading to the abandonment of the tool. This highlights the importance of giving teachers the right to refuse AI powered ed-tech if they are not comfortable with it.

Considering a direct correlation between emotions and learning [ 40 ] recently, ITS have also started focusing on emotional state of students while learning to offer a more contextualized learning experience [ 24 ].

2.5 Popular conferences

To reflect on the increasing interest and activity in the space of AIEd, some of the most popular conferences in AIEd are shown in Table 1 below. Due to the pandemic all these conferences will be available virtually in 2021 as well. The first international workshop on multimodal artificial intelligence in education is being organized at AIEd [74] conference to promote the importance of multimodal data in AIEd.

3 Industry’s focus

In this section, we introduce the industry focus in the area of AIEd by case-studying three levels of companies start-up level, established/large company and mega-players (Amazon, Cisco). These companies represent different levels of the ecosystem (in terms of size).

3.1 Start-ups

There have been a number of ed-tech companies that are leading the AIEd revolution. New funds are also emerging to invest in ed-tech companies and to help ed-tech start-ups in scaling their products. There has been an increase in investor interest [ 21 ]. In 2020 the amount of investment raised by ed-tech companies more than doubled compared to 2019 (according to Techcrunch). This shows another dimension of pandemic’s effect on ed-tech. With an increase in data coming in during the pandemic, it is expected that industry’s focus on AI powered products will increase.

EDUCATE, a leading accelerator focused on ed-tech companies supported by UCL Institute of Education and European Regional Development Fund was formed to bring research and evidence at the centre of product development for ed-tech. This accelerator has supported more than 250 ed-tech companies and 400 entrepreneurs and helped them focus on evidence-informed product development for education.

Number of ed-tech companies are emerging in this space with interesting business models. Third Space Learning offers maths intervention programs for primary and secondary school students. The company aims to provide low-cost quality tuition to support pupils from disadvantaged backgrounds in UK state schools. They have already offered 8,00,000 h of teaching to around 70,000 students, 50% of who were eligible for free meals. Number of mobile apps like Kaizen Languages, Duolingo and Babbel have emerged that help individuals in learning other languages.

3.2 Established players

Pearson is one of the leading educational companies in the world with operations in more than 70 countries and more than 22,000 employees worldwide. They have been making a transition to digital learning and currently generate 66% of their annual revenue from it. According to Pearson, they have built world’s first AI powered calculus tutor called Aida which is publicly available on the App Store. But, its effectiveness in improving students’ calculus skills without any human intervention is still to be seen.

India based ed-tech company known for creating engaging educational content for students raised investment at a ten billion dollar valuation last year [ 70 ]. Century tech is another ed-tech company that is empowering learning through AI. They claim to use neuroscience, learning science and AI to personalize learning and identifying the unique learning pathways for students in 25 countries. They make more than sixty thousand AI powered smart recommendations to learners every day.

Companies like Pearson and Century Tech are building great technology that is impacting learners across the globe. But the usefulness of their acclaimed AI in helping learners from diverse backgrounds, with unique learning needs and completely different contexts is to be proven. As discussed above, teachers play a very important role on how their AI is used by learners. For this, teacher training is vital to fully understand the strengths and weaknesses of these products. It is very important to have an awareness of where these AI products cannot help or can go wrong so teachers and learners know when to avoid relying on them.

In the past few years, the popularity of Massive Online Open Courses (MOOCS) has grown exponentially with the emergence of platforms like Coursera, Udemy, Udacity, LinkedIn Learning and edX [ 5 , 16 , 28 ]. AI can be utilized to develop a better understanding of learner behaviour on MOOCS, produce better content and enhance learning outcomes at scale. Considering these platforms are collecting huge amounts of data, it will be interesting to see the future applications of AI in offering personalized learning and life-long learning solutions to their users [ 81 ].

3.3 Mega-players

Seeing the business potential of AIEd and the kind of impact it can have on the future of humanity, some of the biggest tech companies around the globe are moving into this space. The shift to online education during the pandemic boosted the demand for cloud services. Amazon’s AWS (Amazon Web Services) as a leader in cloud services provider facilitated institutions like Instituto Colombiano para la Evaluacion de la Educacion (ICFES) to scale their online examination service for 70,000 students. Similarly, LSE utilized AWS to scale their online assessments for 2000 students [ 1 , 3 ].

Google’s CEO Sunder Pichai stated that the pandemic offered an incredible opportunity to re-imagine education. Google has launched more than 50 new software tools during the pandemic to facilitate remote learning. Google Classroom which is a part of Google Apps for Education (GAFE) is being widely used by schools around the globe to deliver education. Research shows that it improves class dynamics and helps with learner participation [ 2 , 29 , 62 , 63 , 69 ].

Before moving onto the ethical dimensions of AIEd, it is important to conclude this section by noting an area that is of critical importance to processing industry and services. Aside from these three levels of operation (start-up, medium, and mega companies), there is the question of development of the AIEd infrastructure. As Luckin [41] points out, “True progress will require the development of an AIEd infrastructure. This will not, however, be a single monolithic AIEd system. Instead, it will resemble the marketplace that has been developed for smartphone apps: hundreds and then thousands of individual AIEd components, developed in collaboration with educators, conformed to uniform international data standards, and shared with researchers and developers worldwide. These standards will enable system-level data collation and analysis that help us learn much more about learning itself and how to improve it”.

4 Ethical AIEd

With a number of mishaps in the real world [ 31 , 80 ], ethics in AI has become a real concern for AI researchers and practitioners alike. Within computer science, there is a growing overlap with the broader Digital Ethics [ 19 ] and the ethics and engineering focused on developing Trustworthy AI [ 11 ]. There is a focus on fairness, accountability, transparency and explainability [ 33 , 82 , 83 , 84 ]. Ethics in AI needs to be embedded in the entire development pipeline, from the decision to start collecting data till the point when the machine learning model is deployed in production. From an engineering perspective, Koshiyama et al. [ 35 ] have identified four verticals of algorithmic auditing. These include performance and robustness, bias and discrimination, interpretability and explainability and algorithmic privacy.

In education, ethical AI is crucial to ensure the wellbeing of learners, teachers and other stakeholders involved. There is a lot of work going on in AIEd and AI powered ed-tech tools. With the influx of large amounts of data due to online learning during the pandemic, we will most likely see an increasing number of AI powered ed-tech products. But ethics in AIEd is not a priority for most ed-tech companies and schools. One of the reasons for this is the lack of awareness of relevant stakeholders regarding where AI can go wrong in the context of education. This means that the drawbacks of using AI like discrimination against certain groups due to data deficiencies, stigmatization due to reliance on certain machine learning modelling deficiencies and exploitation of personal data due to lack of awareness can go unnoticed without any accountability.

An AI wrongly predicting that a particular student will not perform very well in end of year exams or might drop out next year can play a very important role in determining that student’s reputation in front of teachers and parents. This reputation will determine how these teachers and parents treat that learner, resulting in a huge psychological impact on that learner, based on this wrong description by an AI tool. One high-profile case of harm was in the use of an algorithm to predict university entry results for students unable to take exams due to the pandemic. The system was shown to be biased against students from poorer backgrounds. Like other sectors where AI is making a huge impact, in AIEd this raises an important ethical question regarding giving students the freedom to opt out of AI powered predictions and automated evaluations.

The ethical implications of AI in education are dependent on the kind of disruption AI is doing in the ed-tech sector. On the one hand, this can be at an individual level for example by recommending wrong learning materials to students, or it can collectively impact relationships between different stakeholders such as how teachers perceive learners’ progress. This can also lead to automation bias and issues of accountability [ 67 ] where teachers begin to blindly rely on AI tools and prefer the tool’s outcomes over their own better judgement, whenever there is a conflict.

Initiatives have been observed in this space. For example, Professor Rose Luckin, professor of learner centered design at University College London along with Sir Anthony Seldon, vice chancellor of the University of Buckingham and Priya Lakhani, founder and CEO of Century Tech founded the Institute of Ethical AI in Education (IEAIEd) [ 72 ] to create awareness and promote the ethical aspects of AI in education. In its interim report, the institute identified seven different requirements for ethical AI to mitigate any kind of risks for learners. This included human agency and oversight to double-check AI’s performance, technical robustness and safety to prevent AI going wrong with new data or being hacked; diversity to ensure similar distribution of different demographics in data and avoid bias; non-discrimination and fairness to prevent anyone from being unfairly treated by AI; privacy and data governance to ensure everyone has the right to control their data; transparency to enhance the understanding of AI products; societal and environmental well-being to ensure that AI is not causing any harm and accountability to ensure that someone takes the responsibility for any wrongdoings of AI. Recently, the institute has also published a framework [ 71 ] for educators, schools and ed-tech companies to help them with the selection of ed-tech products with various ethical considerations in mind, like ethical design, transparency, privacy etc.

With the focus on online learning during the pandemic, and more utilization of AI powered ed-tech tools, risks of AI going wrong have increased significantly for all the stakeholders including ed-tech companies, schools, teachers and learners. A lot more work needs to be done on ethical AI in learning contexts to mitigate these risks, including assessment balancing risks and opportunities.

UNESCO published ‘Beijing Consensus’ on AI and Education that recommended member states to take a number of actions for the smooth and positively impactful integration of AI with education [ 74 ]. International bodies like EU have also recently published a set of draft guidelines under the heading of EU AI Act to ban certain uses of AI and categorize some as ‘high risk’ [ 47 ].

5 Future work

With the focus on online education due to Covid’19 in the past year, it will be consequential to see what AI has to offer for education with vast amounts of data being collected online through Learning Management Systems (LMS) and Massive Online Open Courses (MOOCS).

With this influx of educational data, AI techniques such as reinforcement learning can also be utilized to empower ed-tech. Such algorithms perform best with the large amounts of data that was limited to very few ed-tech companies in 2021. These algorithms have achieved breakthrough performance in multiple domains including games [ 66 ], healthcare [ 14 ] and robotics [ 34 ]. This presents a great opportunity for AI’s applications in education for further enhancing students’ learning outcomes, reducing teachers’ workloads [ 30 ] and making learning personalized [ 64 ], interactive and fun [ 50 , 53 ] for teachers and students.

With a growing number of AI powered ed-tech products in future, there will also be a lot of research on ethical AIEd. The risks of AI going wrong in education and the psychological impact this can have on learners and teachers is huge. Hence, more work needs to be done to ensure robust and safe AI products for all the stakeholders.

This can begin from the ed-tech companies sharing detailed guidelines for using AI powered ed-tech products, particularly specifying when not to rely on them. This includes the detailed documentation of the entire machine learning development pipeline with the assumptions made, data processing approaches used and the processes followed for selecting machine learning models. Regulators can play a very important role in ensuring that certain ethical principles are followed in developing these AI products or there are certain minimum performance thresholds that these products achieve [ 32 ].

6 Conclusion

AIEd promised a lot in its infancy around 3 decades back. However, there are still a number of AI breakthroughs required to see that kind of disruption in education at scale (including basic infrastructure). In the end, the goal of AIEd is not to promote AI, but to support education. In essence, there is only one way to evaluate the impact of AI in Education: through learning outcomes. AIEd for reducing teachers’ workload is a lot more impactful if the reduced workload enables teachers to focus on students’ learning, leading to better learning outcomes.

Cutting edge AI by researchers and companies around the world is not of much use if it is not helping the primary grade student in learning. This problem becomes extremely challenging because every learner is unique with different learning pathways. With the recent developments in AI, particularly reinforcement learning techniques, the future holds exciting possibilities of where AI will take education. For impactful AI in education, learners and teachers always need to be at the epicenter of AI development.

About Amazon.: Helping 7,00,000 students transition to remote learning. https://www.aboutamazon.com/news/community/helping-700-000-students-transition-to-remote-learning (2020)

Al-Maroof, R.A.S., Al-Emran, M.: Students acceptance of google classroom: an exploratory study using PLS–SEM approach. Int. J. Emerg. Technol Learn. (2018). https://doi.org/10.3991/ijet.v13i06.8275

Article   Google Scholar  

Amazon Web Services, Inc. (n.d.).: Amazon Web Services, Inc. https://pages.awscloud.com/whitepaper-emerging-trends-in-education.html (2020)

Baker, R.S.: Stupid tutoring systems, intelligent humans. Int. J. Artif. Intell. Educ. 26 (2), 600–614 (2016)

Baturay, M.H.: An overview of the world of MOOCs. Procedia. Soc. Behav. Sci. 174 , 427–433 (2015)

Baylari, A., Montazer, G.A.: Design a personalized e-learning system based on item response theory and artificial neural network approach. Expert. Syst. Appl. 36 (4), 8013–8021 (2009)

Beck, J., Stern, M., Haugsjaa, E.: Applications of AI in education. Crossroads 3 (1), 11–15 (1996). https://doi.org/10.1016/j.eswa.2008.10.080

Beck, J.E.: Modeling the Student with Reinforcement Learning. Proceedings of the Machine learning for User Modeling Workshop at the Sixth International Conference on User Modeling (1997)

Beck, J.E., Woolf, B.P., Beal, C.R.: ADVISOR: A machine learning architecture for intelligent tutor construction. Proceedings of the 7th National Conference on Artificial Intelligence, New York, ACM, 552–557 (2000)

Boekaerts, M.: Self-regulated learning: where we are today. Int. J. Educ. Res. 31 (6), 445–457 (1999)

Brundage, M., Avin, S., Wang, J., Belfield, H., Krueger, G., Hadfield, G., Maharaj, T.: Toward trustworthy AI development: mechanisms for supporting verifiable claims. arXiv preprint arXiv:2004.07213 (2020)

Bull, S., Kay, J.: Open learner models. In: Nkambou, R., Bourdeau, J., Mizoguchi, R. (eds.) Studies in computational intelligence, pp. 301–322. Springer, Berlin (2010)

Google Scholar  

Cunha-Perez, C., Arevalillo-Herraez, M., Marco-Gimenez, L., Arnau, D.: On incorporating affective support to an intelligent tutoring system: an empirical study. IEEE. R. Iberoamericana. De. Tecnologias. Del. Aprendizaje. 13 (2), 63–69 (2018)

Callaway, E.: “It will change everything”: DeepMind’s AI makes gigantic leap in solving protein structures. Nature. https://www.nature.com/articles/d41586-020-03348-4 . (2020)

Cambridge University Press and Educate Ventures. Shock to the system: lessons from Covid-19 Volume 1: Implications and recommendations. https://www.cambridge.org/pk/files/1616/1349/4545/Shock_to_the_System_Lessons_from_Covid19_Volume_1.pdf (2021). Accessed 12 Apr 2021

Deng, R., Benckendorff, P., Gannaway, D.: Progress and new directions for teaching and learning in MOOCs. Comput. Educ. 129 , 48–60 (2019)

Erümit, A.K., Çetin, İ: Design framework of adaptive intelligent tutoring systems. Educ. Inf. Technol. 25 (5), 4477–4500 (2020)

Fang, Y., Ren, Z., Hu, X., Graesser, A.C.: A meta-analysis of the effectiveness of ALEKS on learning. Educ. Psychol. 39 (10), 1278–1292 (2019)

Floridi, L.: Soft ethics, the governance of the digital and the general data protection regulation. Philos. Trans. R. Soc. A. Math. Phys. Eng. Sci. 376 (2133), 20180081 (2018)

Goldstein, I.J.: The genetic graph: a representation for the evolution of procedural knowledge. Int. J. Man. Mach. Stud. 11 (1), 51–77 (1979)

Goryachikh, S.P., Sozinova, A.A., Grishina, E.N., Nagovitsyna, E.V.: Optimisation of the mechanisms of managing venture investments in the sphere of digital education on the basis of new information and communication technologies: audit and reorganisation. IJEPEE. 13 (6), 587–594 (2020)

Grawemeyer, B., Gutierrez-Santos, S., Holmes, W., Mavrikis, M., Rummel, N., Mazziotti, C., Janning, R.: Talk, tutor, explore, learn: intelligent tutoring and exploration for robust learning, p. 2015. AIED, Madrid (2015)

Hansen, A., Mavrikis, M.: Learning mathematics from multiple representations: two design principles. ICTMT-12, Faro (2015)

Hasan, M.A., Noor, N.F.M., Rahman, S.S.A., Rahman, M.M.: The transition from intelligent to affective tutoring system: a review and open issues. IEEE Access (2020). https://doi.org/10.1109/ACCESS.2020.3036990

Heffernan, N.T., Heffernan, C.L.: The ASSISTments ecosystem: building a platform that brings scientists and teachers together for minimally invasive research on human learning and teaching. Int. J. Artif. Intell. Educ. (2014). https://doi.org/10.1007/s40593-014-0024-x

Heffernan, N.T., Koedinger, K.R.: An intelligent tutoring system incorporating a model of an experienced human tutor. Proceedings of the 6th International Conference on Intelligent Tutoring Systems, 2363, p 596–608, (2002)

Hill, P., Barber, M.: Preparing for a Renaissance in Assessment. Pearson, London (2014)

Hollands, F.M., Tirthali, D.: Why do institutions offer MOOCs? Online Learning 18 (3), 3 (2014)

Iftakhar, S.: Google classroom: what works and how. J. Educ. Soc. Sci. 3 (1), 12–18 (2016)

Iglesias, A., Martínez, P., Aler, R., Fernández, F.: Reinforcement learning of pedagogical policies in adaptive and intelligent educational systems. Knowl. Based. Syst. 22 (4), 266–270 (2009)

Johnson, D.G., Verdicchio, M.: AI, agency and responsibility: the VW fraud case and beyond. Ai. Soc. 34 (3), 639–647 (2019)

Kazim, E., Denny, D.M.T., Koshiyama, A.: AI auditing and impact assessment: according to the UK information commissioner’s office. AI. Ethics. 1 , 1–10 (2021)

Kazim, E., Koshiyama, A.: A High-Level Overview of AI Ethics. SSRN J (2020). https://doi.org/10.2139/ssrn.3609292

Kober, J., Bagnell, J.A., Peters, J.: Reinforcement learning in robotics: a survey. Int. J. Robot. Res. 32 (11), 1238–1274 (2013)

Koshiyama, A., Kazim, E., Treleaven, P., Rai, P., Szpruch, L., Pavey, G., Ahamat, G., Leutner, F., Goebel, R., Knight, A., Adams, J., Hitrova, C., Barnett, J., Nachev, P., Barber, D., Chamorro-Premuzic, T., Klemmer, K., Gregorovic, M., Khan, S., Lomas, E.: Towards algorithm auditing a survey on managing legal ethical and technological risks of AI, ML and associated algorithms. SSRN J (2021). https://doi.org/10.2139/ssrn.3778998

LaPierre, J.: How AI Enhances Collaborative Learning. Filament Games (2018). https://www.filamentgames.com/blog/how-ai-enhances-collaborative-learning/ . Accessed 12 Apr 2021

Luckin, R.: Towards artificial intelligence-based assessment systems. Nat. Hum. Behav. (2017). https://doi.org/10.1038/s41562-016-0028

Luckin, R., du Boulay, B.: Int. J. Artif. Intell. Educ. 26 , 416–430 (2016)

Luckin, R., Holmes, W., Griffiths, M., Pearson, L.: Intelligence Unleashed An argument for AI in Education. https://static.googleusercontent.com/media/edu.google.com/en//pdfs/Intelligence-Unleashed-Publication.pdf (2016)

Barron-Estrada M.L., Zatarain-Cabada, R., Oramas-Bustillos, R., Gonzalez-Hernandez, F.: Sentiment analysis in an affective intelligent tutoring system. Proc. IEEE 17th Int. Conf. Adv. Learn. Technol. (ICALT), Timisoara pp. 394–397 2017.

Ma, W., Adesope, O., Nesbit, J.C., Liu, Q.: Intelligent tutoring systems and learning outcomes: a meta-analysis. J. Educ. Psychol. 106 (4), 901–918 (2014)

Makridakis, S.: The forthcoming Artificial Intelligence (AI) revolution: Its impact on society and firms. Futures 90 , 46–60 (2017)

Mavrikis, M.: Int. J. Artif. Intell. Tools. 19 , 733–753 (2010)

Merrill, D.C., Reiser, B.J., Ranney, M., Trafton, J.G.: Effective tutoring techniques: a comparison of human tutors and intelligent tutoring systems. J. Learn. Sci. 2 (3), 277–305 (1992)

Moeini, A.: Theorising Evidence-Informed Learning Technology Enterprises: A Participatory Design-Based Research Approach. Doctoral dissertation, UCL University College London, London, (2020)

Mohamed, H., Lamia, M.: Implementing flipped classroom that used an intelligent tutoring system into learning process. Comput. Educ. 124 , 62–76 (2018). https://doi.org/10.1016/j.compedu.2018.05.011

Mueller, B.: The Artificial Intelligence Act: A Quick Explainer. [online] Center for Data Innovation (2021). https://datainnovation.org/2021/05/the-artificial-intelligence-act-a-quick-explainer/ . Accessed 12 Apr 2021

Murray, M.C., Pérez, J.: Informing and performing: A study comparing adaptive learning to traditional learning. Inform. Sci. J. 18 , 111–125 (2015)

Oudeyer, P-Y.: Computational Theories of Curiosity-Driven Learning. https://arxiv.org/pdf/1802.10546.pdf (2018)

Park, H.W., Grover, I., Spaulding, S., Gomez, L., Breazeal, C.: A model-free affective reinforcement learning approach to personalization of an autonomous social robot companion for early literacy education. AAAI. 33 (1), 687–694 (2019)

Resnick, M., Robinson, K.: Lifelong kindergarten: cultivating creativity through projects, passion, peers, and play. MIT press, Cambridge (2017)

Book   Google Scholar  

Rodríguez-Triana, M.J., Prieto, L.P., Martínez-Monés, A., Asensio-Pérez, J.I. and Dimitriadis, Y.: The teacher in the loop: Customizing multimodal learning analytics for blended learning. In Proceedings of the 8th international conference on learning analytics and knowledge. pp 417–426 (2018)

Rowe, J.P., Lester, J.C.: Improving student problem solving in narrative-centered learning environments: a modular reinforcement learning framework. In International Conference on Artificial Intelligence in Education. pp. 419–428. Springer, Cham (2015)

Russell, S.J., Norvig, P., Davis, E.: Artificial intelligence: a modern approach. Prentice Hall, Upper Saddle River (2010)

MATH   Google Scholar  

Jiménez, S., Juárez-Ramírez, R., Castillo, V.H., Licea, G., Ramírez-Noriega, A., Inzunza, S.: A feedback system to provide affective support to students. Comput. Appl. Eng. Educ. 26 (3), 473–483 (2018)

Sadler, D.R.: Formative assessment in the design of instructional systems. Instr. Sci. 18 , 119–144 (1989)

Samarakou, M., Fylladitakis, E., Prentakis, P., Athineos, S.: Implementation of artificial intelligence assessment in engineering laboratory education. https://files.eric.ed.gov/fulltext/ED557263.pdf (2014). Accessed 24 Feb 2021

Segal, A., Hindi, S., Prusak, N., Swidan, O., Livni, A., Palatnic, A., Schwarz, B.: Keeping the teacher in the loop: Technologies for monitoring group learning in real-time. In International Conference on Artificial Intelligence in Education. pp. 64–76. Springer, Cham (2017)

Self, J. A. (1990). Theoretical foundations of intelligent tutoring systems. J. Artif. Intell

Self, J.A.: The defining characteristics of intelligent tutoring systems research: ITSs care, precisely. IJAIEd. 10 , 350–364 (1998)

Selwood, I., Pilkington, R.: Teacher workload: using ICT to release time to teach. Educ. Rev. 57 (2), 163–174 (2005)

Shaharanee, I.N.M., Jamil, J.M. and Rodzi, S.S.M.:æ Google classroom as a tool for active learning. AIP Conference Proceedings, 1761 (1), pp. 020069, AIP Publishing LLC, College Park (2016)

Shaharanee, I.N.M., Jamil, J.M., Rodzi, S.S.M.: The application of Google Classroom as a tool for teaching and learning. J. Telecommun. Electron. Comp. Eng. 8 (10), 5–8 (2016)

Shawky, D., Badawi, A.: Towards a personalized learning experience using reinforcement learning. In: Hassanien, A.E. (ed.) Machine learning paradigms Theory and application, pp. 169–187. Springer (2019)

Shute, V.J. (1991). Rose garden promises of intelligent tutoring systems: blossom or thorn.  NASA, Lyndon B. Johnson Space Center, Fourth Annual Workshop on Space Operations Applications and Research (SOAR 90) . Available at: https://ntrs.nasa.gov/citations/19910011382 . Accessed 4 July 2021

Silver, D., Huang, A., Maddison, C.J., Guez, A., Sifre, L., Van Den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S.: Mastering the game of Go with deep neural networks and tree search. Nature 529 (7587), 484–489 (2016)

Skitka, L.J., Mosier, K., Burdick, M.D.: Accountability and automation bias. Int. J. Hum. Comput. Stud. 52 (4), 701–717 (2000)

Steenbergen-Hu, S., Cooper, H.: A meta-analysis of the effectiveness of intelligent tutoring systems on K–12 students’ mathematical learning. J. Educ. Psychol. 105 (4), 970–987 (2013)

Sudarsana, I.K., Putra, I.B., Astawa, I.N.T., Yogantara, I.W.L.: The use of google classroom in the learning process. J. Phys. Conf. Ser 1175 (1), 012165 (2019)

TechCrunch. Indian education startup Byju’s is fundraising at a $10B valuation. https://techcrunch.com/2020/05/01/indian-education-startup-byjus-is-fundraising-at-a-10b-valuation/ (2020). Accessed 12 Apr 2021

The Institute for Ethical AI in Education The Ethical Framework for AI in Education (IEAIED). https://fb77c667c4d6e21c1e06.b-cdn.net/wp-content/uploads/2021/03/The-Ethical-Framework-for-AI-in-Education-Institute-for-Ethical-AI-in-Education-Final-Report.pdf (2021). Accessed 12 Apr 2021

The Institute for Ethical AI in Education The Ethical Framework for AI in Education (n.d.). Available at: https://www.buckingham.ac.uk/wp-content/uploads/2021/03/The-Institute-for-Ethical-AI-in-Education-The-Ethical-Framework-for-AI-in-Education.pdf . Accessed 4 July 2021

Tisseron, S., Tordo, F., Baddoura, R.: Testing Empathy with Robots: a model in four dimensions and sixteen ítems. Int. J. Soc. Robot. 7 (1), 97–102 (2015)

UNESCO. Artificial intelligence in education. UNESCO. https://en.unesco.org/artificial-intelligence/education . (2019). Accessed 12 Apr 2021

Utterberg Modén, M., Tallvid, M., Lundin, J., Lindström, B.: Intelligent Tutoring Systems: Why Teachers Abandoned a Technology Aimed at Automating Teaching Processes. In: Proceedings of the 54th Hawaii International Conference on System Sciences, Maui, p. 1538 (2021)

van der Spoel, I., Noroozi, O., Schuurink, E., van Ginkel, S.: Teachers’ online teaching expectations and experiences during the Covid19-pandemic in the Netherlands. Eur. J. Teach. Educ. 43 (4), 623–638 (2020)

Weller, M.: Twenty years of EdTech. Educa. Rev. Online. 53 (4), 34–48 (2018)

Wenger, E.: Artificial intelligence and tutoring systems. Morgan Kauffman, Los Altos (1987)

World Economic Forum and The Boston Consulting Group. New vision for education unlocking the potential of technology industry agenda prepared in collaboration with the Boston consulting group. http://www3.weforum.org/docs/WEFUSA_NewVisionforEducation_Report2015.pdf (2015). Accessed 12 Apr 2021

Yampolskiy, R.V., Spellchecker, M.S.: Artificial intelligence safety and cybersecurity: a timeline of AI failures. arXiv:1610.07997 (2016)

Yu, H., Miao, C., Leung, C., White, T.J.: Towards AI-powered personalization in MOOC learning. Npj. Sci. Learn. 2 (1), 1–5 (2017)

Yu, H., Shen, Z., Miao, C., Leung, C., Lesser, V.R., Yang, Q.: Building ethics into artificial intelligence. arXiv:1812.02953 (2018)

Zemel, R., Wu Y., Swersky, K., Pitassi, T., Dwork, C.: Learning fair representations. In: International Conference on Machine Learning, pp. 325–333 (2013)

Zhang, Y., Liao, Q.V., Bellamy, R.K.E.: Effect of confidence and explanation on accuracy and trust calibration in AI-assisted decision making. In: Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency. arXiv:2001.02114 (2020)

Zimmerman, B.J., Schunk, D.H.: Handbook of Self-Regulation of Learning and Performance. Routledge, Oxfordshire (2011)

Download references

Author information

Authors and affiliations.

Artificial Intelligence at University College, London, UK

Muhammad Ali Chaudhry

Department of Computer Science, University College, London, UK

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Muhammad Ali Chaudhry .

Additional information

Publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Chaudhry, M.A., Kazim, E. Artificial Intelligence in Education (AIEd): a high-level academic and industry note 2021. AI Ethics 2 , 157–165 (2022). https://doi.org/10.1007/s43681-021-00074-z

Download citation

Received : 25 April 2021

Accepted : 17 June 2021

Published : 07 July 2021

Issue Date : February 2022

DOI : https://doi.org/10.1007/s43681-021-00074-z

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

  • Artificial Intelligence
  • Machine learning
  • Learning science
  • Artificial Intelligence in Education (AIEd)
  • Intelligent Tutoring Systems (ITS)
  • Find a journal
  • Publish with us
  • Track your research

IEEE Account

  • Change Username/Password
  • Update Address

Purchase Details

  • Payment Options
  • Order History
  • View Purchased Documents

Profile Information

  • Communications Preferences
  • Profession and Education
  • Technical Interests
  • US & Canada: +1 800 678 4333
  • Worldwide: +1 732 981 0060
  • Contact & Support
  • About IEEE Xplore
  • Accessibility
  • Terms of Use
  • Nondiscrimination Policy
  • Privacy & Opting Out of Cookies

A not-for-profit organization, IEEE is the world's largest technical professional organization dedicated to advancing technology for the benefit of humanity. © Copyright 2024 IEEE - All rights reserved. Use of this web site signifies your agreement to the terms and conditions.

research paper based on artificial intelligence

The Journal of Artificial Intelligence Research (JAIR) is dedicated to the rapid dissemination of important research results to the global artificial intelligence (AI) community. The journal’s scope encompasses all areas of AI, including agents and multi-agent systems, automated reasoning, constraint processing and search, knowledge representation, machine learning, natural language, planning and scheduling, robotics and vision, and uncertainty in AI.

Current Issue

Vol. 79 (2024)

Published: 2024-01-10

Can Fairness be Automated? Guidelines and Opportunities for Fairness-aware AutoML

Practical and parallelizable algorithms for non-monotone submodular maximization with size constraint, exploring the tradeoff between system profit and income equality among ride-hailing drivers, on mitigating the utility-loss in differentially private learning: a new perspective by a geometrically inspired kernel approach, an algorithm with improved complexity for pebble motion/multi-agent path finding on trees, weighted, circular and semi-algebraic proofs, reinforcement learning for generative ai: state of the art, opportunities and open research challenges, human-in-the-loop reinforcement learning: a survey and position on requirements, challenges, and opportunities, boolean observation games, detecting change intervals with isolation distributional kernel, query-driven qualitative constraint acquisition, visually grounded language learning: a review of language games, datasets, tasks, and models, right place, right time: proactive multi-robot task allocation under spatiotemporal uncertainty, principles and their computational consequences for argumentation frameworks with collective attacks, the ai race: why current neural network-based architectures are a poor basis for artificial general intelligence, undesirable biases in nlp: addressing challenges of measurement.

research paper based on artificial intelligence

Employing nano-enabled artificial intelligence (AI)-based smart technologies for prediction, screening, and detection of cancer

ORCID logo

* Corresponding authors

a Quantum Materials and Devices Unit, Institute of Nano Science and Technology, Mohali, Punjab 140306, India E-mail: [email protected] , [email protected]

b Biological Science, St. John's University, New York, NY 10301, United States

c NanoBioTech Laboratory, Department of Environmental Engineering, Florida Polytechnic University, Lakeland, Florida 33805, USA

d Electrical and Computer Engineering, Florida International University, Miami, FL 33199, USA

Cancer has been classified as a diverse illness with a wide range of subgroups. Its early identification and prognosis, which have become a requirement of cancer research, are essential for clinical treatment. Patients have already benefited greatly from the use of artificial intelligence (AI), machine learning (ML), and deep learning (DL) algorithms in the field of healthcare. AI simulates and combines data, pre-programmed rules, and knowledge to produce predictions. Data are used to improve efficiency across several pursuits and tasks through the art of ML. DL is a larger family of ML methods based on representational learning and simulated neural networks. Support vector machines, convulsion neural networks, and artificial neural networks, among others, have been widely used in cancer research to construct prediction models that enable precise and effective decision-making. Although using these innovative methods can enhance our comprehension of how cancer progresses, further validation is required before these techniques can be used in routine clinical practice. We cover contemporary methods used in the modelling of cancer development in this article. The presented prediction models are built using a variety of guided ML approaches, as well as numerous input attributes and data collections. Early identification and cost-effective detection of cancer's progression are equally necessary for successful treatment of the disease. Smart material-based detection techniques can give end consumers a portable, affordable instrument to easily detect and monitor their health issues without the need for specialized knowledge. Owing to their cost-effectiveness, excellent sensitivity, multimodal detection capacity, and miniaturization aptitude, two-dimensional (2D) materials have a lot of prospects for clinical examination of various compounds as well as cancer biomarkers. The effectiveness of traditional devices is moving faster towards more useful techniques thanks to developments in 2D material-based biosensors/sensors. The most current developments in the design of 2D material-based biosensors/sensors—the next wave of cancer screening instruments—are also outlined in this article.

Graphical abstract: Employing nano-enabled artificial intelligence (AI)-based smart technologies for prediction, screening, and detection of cancer

  • This article is part of the themed collections: Recent Review Articles and Nanoscale 2024 Emerging Investigators

Article information

Download citation, permissions.

research paper based on artificial intelligence

V. Chugh, A. Basu, A. Kaushik, Manshu, S. Bhansali and A. K. Basu, Nanoscale , 2024, Advance Article , DOI: 10.1039/D3NR05648A

To request permission to reproduce material from this article, please go to the Copyright Clearance Center request page .

If you are an author contributing to an RSC publication, you do not need to request permission provided correct acknowledgement is given.

If you are the author of this article, you do not need to request permission to reproduce figures and diagrams provided correct acknowledgement is given. If you want to reproduce the whole article in a third-party publication (excluding your thesis/dissertation for which permission is not required) please go to the Copyright Clearance Center request page .

Read more about how to correctly acknowledge RSC content .

Social activity

Search articles by author.

This article has not yet been cited.

Advertisements

Thank you for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

  • View all journals
  • My Account Login
  • Explore content
  • About the journal
  • Publish with us
  • Sign up for alerts
  • Open access
  • Published: 21 February 2024

Avoiding fusion plasma tearing instability with deep reinforcement learning

  • Jaemin Seo   ORCID: orcid.org/0000-0003-0635-0282 1 , 2 ,
  • SangKyeun Kim 1 , 3 ,
  • Azarakhsh Jalalvand 1 ,
  • Rory Conlin 1 , 3 ,
  • Andrew Rothstein 1 ,
  • Joseph Abbate 3 , 4 ,
  • Keith Erickson 3 ,
  • Josiah Wai 1 ,
  • Ricardo Shousha 1 , 3 &
  • Egemen Kolemen   ORCID: orcid.org/0000-0003-4212-3247 1 , 3  

Nature volume  626 ,  pages 746–751 ( 2024 ) Cite this article

6610 Accesses

245 Altmetric

Metrics details

  • Computational science
  • Information theory and computation
  • Magnetically confined plasmas
  • Mechanical engineering
  • Nuclear fusion and fission

For stable and efficient fusion energy production using a tokamak reactor, it is essential to maintain a high-pressure hydrogenic plasma without plasma disruption. Therefore, it is necessary to actively control the tokamak based on the observed plasma state, to manoeuvre high-pressure plasma while avoiding tearing instability, the leading cause of disruptions. This presents an obstacle-avoidance problem for which artificial intelligence based on reinforcement learning has recently shown remarkable performance 1 , 2 , 3 , 4 . However, the obstacle here, the tearing instability, is difficult to forecast and is highly prone to terminating plasma operations, especially in the ITER baseline scenario. Previously, we developed a multimodal dynamic model that estimates the likelihood of future tearing instability based on signals from multiple diagnostics and actuators 5 . Here we harness this dynamic model as a training environment for reinforcement-learning artificial intelligence, facilitating automated instability prevention. We demonstrate artificial intelligence control to lower the possibility of disruptive tearing instabilities in DIII-D 6 , the largest magnetic fusion facility in the United States. The controller maintained the tearing likelihood under a given threshold, even under relatively unfavourable conditions of low safety factor and low torque. In particular, it allowed the plasma to actively track the stable path within the time-varying operational space while maintaining H-mode performance, which was challenging with traditional preprogrammed control. This controller paves the path to developing stable high-performance operational scenarios for future use in ITER.

As the demand for energy and the need for carbon neutrality continue to grow, nuclear fusion is rapidly emerging as a promising energy source in the near future due to its potential for zero-carbon power generation, without creating high-level waste. Recently, the nuclear fusion experiment accompanied by 192 lasers at the National Ignition Facility successfully produced more energy than the injected energy, demonstrating the feasibility of net energy production 7 . Tokamaks, the most studied concept for the first fusion reactor, have also achieved remarkable milestones: The Korea Superconducting Tokamak Advanced Research sustained plasma at ion temperatures hotter than 100 million kelvin for 30 seconds 8 , a plasma remained in a steady state for 1,000 seconds in the Experimental Advanced Superconducting Tokamak 9 , and the Joint European Torus broke the world record by producing 59 megajoules of fusion energy for 5 seconds 10 , 11 . ITER, the world’s largest science project with the collaboration of 35 nations, is under construction for the demonstration of a tokamak reactor 12 .

Although fusion experiments in tokamaks have achieved remarkable success, there still remain several obstacles that we must resolve. Plasma disruption is one of the most critical issues to be solved for the successful long-pulse operation of ITER 13 . Even a few plasma disruption events can induce irreversible damage to the plasma-facing components in ITER. Recently, techniques for predicting disruption using artificial intelligence (AI) have been demonstrated in multiple tokamaks 14 , 15 , and mitigation of the damage during disruption is being studied 16 , 17 . Tearing instability, the most dominant cause of plasma disruption 18 , especially in the ITER baseline scenario 19 , is a phenomenon where the magnetic flux surface breaks due to finite plasma resistivity at rational surfaces of safety factor q  =  m / n . Here, m and n are the poloidal and toroidal mode numbers, respectively. In modern tokamaks, the plasma pressure is often limited by the onset of neoclassical tearing instability because the perturbation of pressure-driven (so-called bootstrap) current becomes a seed for it 20 . Research on the evolution and suppression of existing tearing instabilities using actuators has been widely conducted 21 , 22 , 23 , 24 , 25 , 26 , 27 . However, the tearing instability induces unrecoverable energy loss and often leads to disruption before being suppressed in the ITER baseline condition, where the edge safety factor ( q 95 ) and plasma rotation are low 19 . Therefore, we need to ‘avoid’ the onset of tearing instability, not suppress it after it appears. To avoid its occurrence, physics research is also underway to investigate the onset cause or seed of instability 28 , 29 , 30 . However, calculating tearing stability requires massive computational simulations based on resistive magnetohydrodynamics or gyrokinetics, which are not suitable for real-time stability prediction and control during experiments. This suggests the need for AI-accelerated real-time instability-avoidance techniques.

The deep reinforcement learning (RL) technique has shown remarkable performance in nonlinear, high-dimensional actuation problems 1 . Moreover, it has shown notable advantages in avoidance control problems 2 , which is essentially similar to the objective of this work. Recently, RL has been applied to tokamak control and optimization, showing promising achievements 3 , 4 , 31 , 32 , 33 , 34 , 35 . The RL algorithm optimizes the actor model based on a deep neural network (DNN), and the actor model gradually learns the action policy leading to higher rewards in a given environment. By specifically designing the reward function, we can train the actor model to actively control the tokamak to pursue a high-pressure plasma while keeping the tearing possibility low. An essential component of RL is the training environment, which can interact with the actor model by responding to its action. For the training environment, we employ a dynamic model that predicts future plasma pressure and tearing likelihood (so-called tearability) developed in ref. 5 . In this work, we develop an AI controller that adaptively controls actuators to pursue high plasma pressure while maintaining low tearability, based on observed plasma profiles. The overall architecture of this tearing-avoidance system is depicted in Fig. 1 .

figure 1

a , The selected diagnostic systems used in this work: magnetics, Thomson scattering (TS) and charge-exchange recombination (CER) spectroscopy. The possible tearing instability of m / n  = 2/1 is shown in orange. b , The heating, current drive and control actuators used in this work. c , Schematic description of the tearing-avoidance control, including preprocessing, high-level control by a DNN and low-level control by a PCS. d , The AI controller based on the DNN.

Figure 1a,b shows an example plasma in DIII-D and selected diagnostics and actuators for this work. A possible tearing instability of m / n  = 2/1 at the flux surface of q  = 2 is also illustrated. Figure 1c shows the tearing-avoidance control system, which maps the measurement signals and the desired actuator commands. The signals from different diagnostics have different dimensions and spatial resolutions, and the availability and target positions of each channel vary depending on the discharge condition. Therefore, the measured signals are preprocessed into structured data of the same dimension and spatial resolution using the profile reconstruction 36 , 37 , 38 and equilibrium fitting (EFIT) 39 before being fed into the DNN model. The DNN-based AI controller (Fig. 1d ) determines the high-level control commands of the total beam power and plasma shape based on the trained control policy. Its training using RL is described in the following section. The plasma control system (PCS) algorithm calculates the low-level control signals of the magnetic coils and the powers of individual beams to satisfy the high-level AI controls, as well as user-prescribed constraints. In our experiments, we constrain q 95 and total beam torque in the PCS to maintain the ITER baseline-similar condition where tearing instability is crucial.

RL design for tearing-avoidance control

For efficient fusion power generation, it is essential to maintain high plasma pressure without disruptive instability. However, as external heating like neutral beams increases the plasma pressure, a stability limit is eventually reached, as shown by the black lines in Fig. 2a , beyond which the tearing instability is excited. The instability can induce plasma disruption shortly, as shown in Fig. 2b,c . Moreover, this stability limit varies depending on the plasma state, and lowering the pressure can also cause instability under certain conditions 19 . As depicted by the blue lines in Fig. 2 , the actuators can be actively controlled depending on the plasma state to pursue high plasma pressure without crossing the onset of instability.

figure 2

a , The time evolution of actuators with (blue) and without (black) the AI control. Possible tearing stability limits are indicated in red. b , The tearability expected by actuators' control. c , The normalized plasma pressure expected by actuators' control. d , The expected plasma evolution by the desired AI control in parametric space.

This is a typical obstacle-avoidance problem, where the obstacle here has a high potential to terminate the operation immediately. We need to control the tokamak to guide the plasma along a narrow acceptable path where the pressure is high enough and the stability limit is not exceeded. To train the actor model for this goal with RL, we designed the reward function, R , to evaluate how high pressure the plasma is under tolerable tearability, as shown in equation ( 1 ). β N represents the normalized plasma pressure, T is the tearability and k is the prescribed threshold. Here, β N and T are the predictions after 25 ms resulting from the action of the AI controller. The prediction of future β N and T using a dynamic model is described in more detail in Methods . The threshold k is set to 0.2, 0.5 or 0.7 in this work. If the predicted tearability is below a given threshold, the actor receives a positive reward based on the attained plasma pressure, and it receives a negative reward otherwise.

To obtain a higher reward, defined in equation ( 1 ), the actor should first increase β N through its control actions. However, higher β N tends to make the plasma unstable, causing the tearability ( T ) to exceed the threshold ( k ) at some point, which in turn reduces the reward. We note that the reward shows a steep change when T exceeds k , like a binary penalty. This leads the actor model to prioritize maintaining T below k over increasing β N . After sufficient training with RL, the actor can determine the control actions that pursue high plasma pressure while keeping the tearability below the given threshold. This control policy enables the tokamak operation to follow a narrow desired path during a discharge, as illustrated in Fig. 2d . It is noted that the reward contour surface in Fig. 2d is a simplified representation for illustrative purposes, while the actual reward contour according to equation ( 1 ) has a sharp bifurcation near the tearing onset.

The action variables controlled by AI are set as the total beam power and the plasma triangularity. Although there are other controllable actuators through the PCS, such as the beam torque, plasma current or plasma elongation, they strongly affect q 95 and the plasma rotation. Thus, for the purpose of maintaining the ITER baseline-similar condition of q 95  ≈ 3 and beam torque ≤1 Nm, these other actuators were fixed during the experiments.

The observation variables are set as one-dimensional kinetic and magnetic profiles mapped in a magnetic flux coordinate because the tearing onset strongly depends on their spatial information and gradients 19 . Specifically, the actor observes profiles of the electron density, electron temperature, ion rotation, safety factor and plasma pressure. An example set of observation profiles is shown in Fig. 3a .

figure 3

a , The observation of the AI controller; the preprocessed profiles of electron density, electron temperature, ion rotation, safety factor and plasma pressure. b , The time traces of discharges 193266 (stable reference), 193273 (unstable reference) and 193280. Discharge 193280 is the AI-controlled one. c , The low-level coil current control by the PCS and the plasma boundary variation. Scaled currents of poloidal field (PF) coils are shown in colour. d , The low-level individual beam power control by the PCS. e , The estimated tearability for discharges 193273 and 193280.

Tearing-avoidance control in DIII-D

An example of plasma disruption due to tearing instability is depicted by the black lines (discharge 193273) in Fig. 3b . In discharge 193273, a traditional feedback control (not AI control) was applied to maintain β N  = 2.3. However, at t  = 2.6 s, a large tearing instability occurred, as shown in the fourth row of Fig. 3b . This led to unrecoverable degradation of β N , eventually resulting in a disruption at t  = 3.1 s. This indicates that the tearing onset boundary is crossed at some point before t  = 2.6 s. Figure 3e depicts the post-experiment tearability prediction for this discharge. This post-analysis reveals that the tearing event could have been forecasted as early as 200 ms beforehand, providing sufficient time to lower tearability via appropriate control. As the model predicts the onset of tearing instability, not classifies whether the current state is tearing or not, the tearability decreases back to 0 after the onset passes ( t  > 2.7 s). The yellow line (discharge 193266) in Fig. 3b , which targets β N  = 1.7 under traditional control, represents a stable example that could roughly be considered as a conservative bound for tearing stability.

In discharge 193280 (the blue lines in Fig. 3b ), beam power and plasma triangularity were adaptively controlled via AI. Here the AI controller was trained to ensure that the predicted tearability does not exceed 0.5 ( k  = 0.5 in equation ( 1 )). As shown in the second and third rows of Fig. 3b , the AI controller actively adjusts the two actuators according to the time-evolving plasma state. Other controllable parameters were kept fixed during discharge to constrain q 95  ≈ 3 and beam torque ≤1 Nm. At each time point, the AI controller observes the plasma profiles and determines control commands for beam power and triangularity. The PCS algorithm receives these high-level commands and derives low-level actuations, such as magnetic coil currents and the individual powers of the eight beams 39 , 40 , 41 . The coil currents and resulting plasma shape at each phase are shown in Fig. 3c and the individual beam power controls are shown in Fig. 3d .

The blue line in Fig. 3e , a post-experiment estimation for the AI-controlled discharge (193280), shows that the estimated tearability is maintained just below the given threshold until the end, reflecting the exact intention in equation ( 1 ). This experiment demonstrated the ability to achieve lower tearability than the traditional control discharge 193273, and higher time-integrated performance than 193266, through adaptive and active control via AI.

The control policy of a trained actor model can vary depending on the threshold ( k ) of the reward function equation ( 1 ) during the RL training. As the tearability threshold for receiving negative rewards increases, the control policy becomes less conservative. The controller trained with a higher threshold is willing to tolerate higher tearability while pushing β N .

Figure 4a shows three experiments conducted by controllers of different threshold values. Discharges 193277 (grey), 193280 (blue) and 193281 (red) correspond to threshold values of 0.2, 0.5 and 0.7, respectively. In the cases of k  = 0.5 and k  = 0.7, the plasma is sustained without disruptive instability until the preprogrammed end of the flat top. Figure 4b–d shows the post-calculated tearability for the three discharges. The background contour colour in each graph represents the predicted tearability for possible beam powers at each time point, and the actual beam power is indicated by the black line. The dashed lines correspond to the tearability contour lines for each threshold (0.2, 0.5 or 0.7).

figure 4

a , The time traces of discharges with different thresholds; 193277 ( k  = 0.2), 193280 ( k  = 0.5) and 193281 ( k  = 0.7). b – d , The actual beam power and the contour of the predicted tearability for possible beam powers in the three discharges 193277 ( b ), 193280 ( c ) and 193281 ( d ).

Different threshold values result in different characteristics during the AI control in the experiments. In the early phase ( t  < 3.5 s), the high-threshold controller ( k  = 0.7) tends to push β N harder, as shown in the last row of Fig. 4a . However, this leads to putting the plasma in a more unstable region and accepting higher tearability around 0.7 after t  = 3.5 s, and the increased tearability does not decrease afterwards. In contrast, the low-threshold controller ( k  = 0.2) is overly conservative and suppresses the possibility of instability too much in the early phase. The AI control maintained a very low tearability of less than 0.2 until t  = 5 s, but a large instability, difficult to be avoided, suddenly occurred at t  = 5.5 s. As revealed in the post-analysis (Fig. 4b ), the tearing prediction model could forecast the instability 300 ms before the disruption, and the controller also attempted to further reduce the beam power accordingly. However, as the beam power had already reached its prescribed lower bound, it could not be lowered further, ultimately failing to avoid the instability. The lower bound of the beam power was prescribed to prevent L-mode back transition, independent of the RL control, and this was not considered during the training of the controller. As k  = 0.2 is a conservative setting, the controller often attempts to reduce the beam power, which frequently hits the lower bound. As a result, the control interference due to the preset lower bound led to the failure of tearing avoidance. In contrast, the controller with a moderate threshold ( k  = 0.5) sustains the plasma until the end of the flat top and eventually recovers β N again. Therefore, an optimal threshold value is required to maintain stable plasma for a long time. In Fig. 4c , the AI controller of k  = 0.5 actively tries to avoid touching the threshold through proactive control before the instability warning. Because the reward in equation ( 1 ) is computed using the tearability 25 ms after the controller’s action at each time point, the trained controller takes actions tens of milliseconds before a warning occurs.

We present a technique for avoiding disruptive tearing instability in a tokamak using the RL method. The AI-based tearing-avoidance system actively controls the beam power and the plasma triangularity to maintain the possibility of future tearing-instability occurrence at a low level. This enabled maintaining the tearability below the threshold under the low- q 95 and low-torque conditions in DIII-D. In addition, our controller has demonstrated the ability to robustly avoid tearing instability not only in a specific experimental condition like the ITER baseline condition but also in other operational environments and even in accidental cases, which is further discussed in Methods .

Our work is a proof-of-concept study on tearing avoidance using RL and is still in the early stages of fine-tuning. For more useful applications, further experiments and fine-tuning are required. Nonetheless, this work demonstrates the capability that RL could be applied to real-time control of core plasma physics, as well as plasma boundary control shown in ref. 3 . We also note that this demonstration is a successful extension of machine-learning capability in the fusion area, bringing insight and a path to developing the integrated control for high-performance operational scenarios in future tokamak devices, beyond the single instability control. There are further potential applications of the tearing-avoidance control developed in this work. For example, this algorithm can be combined with the plasma profile prediction system 42 or physics information 43 , which enables optimizing the entire discharge through combined autoregressive prediction of the plasma state and desired actuator control. In addition, by sustaining plasmas without disruption under extreme conditions, we can discover phenomena such as a new kind of self-generated current 44 , which may help us to achieve efficient fusion energy harvesting.

The DIII-D National Fusion Facility, located at General Atomics in San Diego, USA, is a leading research facility dedicated to advancing the field of fusion energy through experimental and theoretical research. The facility is home to the DIII-D tokamak, which is the largest and most advanced magnetic fusion device in the United States. The major and minor radii of DIII-D are 1.67 m and 0.67 m, respectively. The toroidal magnetic field can reach up to 2.2 T, the plasma current is up to 2.0 MA and the external heating power is up to 23 MW. DIII-D is equipped with high-resolution real-time plasma diagnostic systems, including a Thomson scattering system 45 , charge-exchange recombination 46 spectroscopy and magnetohydrodynamics reconstruction by EFIT 37 , 39 . These diagnostic tools allow for the real-time profiling of electron density, electron temperature, ion temperature, ion rotation, pressure, current density and safety factor. In addition, DIII-D can perform flexible total beam power and torque control through reliable high-frequency modulation of eight different neutral beams in different directions. Therefore, DIII-D is an optimal experimental device for verifying and utilizing our AI controller that observes the plasma state and manipulates the actuators in real time.

Plasma control system

One of the unique features of the DIII-D tokamak is its advanced PCS 47 , which allows researchers to precisely control and manipulate the plasma in real time. This enables researchers to study the behaviour of the plasma under a wide range of conditions and to test ideas for controlling and stabilizing the plasma. The PCS consists of a hierarchical structure of real-time controllers, from the magnetic control system (low-level control) to the profile control system (high-level control). Our tearing-avoidance algorithm is also implemented in this hierarchical structure of the DIII-D PCS and is integrated with the existing lower-level controllers, such as the plasma boundary control algorithm 39 , 41 and the individual beam control algorithm 40 .

Tearing instability

Magnetic reconnection refers to the phenomenon in magnetized plasmas where the magnetic-field line is torn and reconnected owing to the diffusion of magnetic flux ( ψ ) by plasma resistivity. This magnetic reconnection is a ubiquitous event occurring in diverse environments such as the solar atmosphere, the Earth’s magnetosphere, plasma thrusters and laboratory plasmas like tokamaks. In nested magnetic-field structures in tokamaks, magnetic reconnection at surfaces where q becomes a rational number leads to the formation of separated field lines creating magnetic islands. When these islands grow and become unstable, it is termed tearing instability. The growth rate of the tearing instability classically depends on the tearing stability index, Δ ′, shown in equation ( 2 ).

where x is the radial deviation from the rational surface. When Δ ′ is positive, the magnetic topology becomes unstable, allowing (classical) tearing instability to develop. However, even when Δ ′ is negative (classical tearing instability does not grow), ‘neoclassical’ tearing instability can arise due to the effects of geometry or the drift of charged particles, which can amplify seed perturbations. Subsequently, the altered magnetic topology can either saturate, unable to grow further 48 , 49 , or can couple with other magnetohydrodynamic events or plasma turbulence 50 , 51 , 52 , 53 . Understanding and controlling these tearing instabilities is paramount for achieving stable and sustainable fusion reactions in a tokamak 54 .

ITER baseline scenario

The ITER baseline scenario (IBS) is an operational condition designed for ITER to achieve fusion power of P fusion  = 500 MW and a fusion gain of Q  ≡  P fusion / P external  = 10 for a duration of longer than 300 s (ref. 12 ). Compared with present tokamak experiments, the IBS condition is notable for its considerably low edge safety factor ( q 95  ≈ 3) and toroidal torque. With the PCS, DIII-D has a reliable capability to access this IBS condition compared with other devices; however, it has been observed that many of the IBS experiments are terminated by disruptive tearing instabilities 19 . This is because the tearing instability at the q  = 2 surface appears too close to the wall when q 95 is low, and it easily locks to the wall, leading to disruption when the plasma rotation frequency is low. Therefore, in this study, we conducted experiments to test the AI tearability controller under the conditions of q 95  ≈ 3 and low toroidal torque (≤1 Nm), where the disruptive tearing instability is easy to be excited.

However, in addition to the IBS where the tearing instability is a critical issue, there are other scenarios, such as hybrid and non-inductive scenarios for ITER 12 . These different scenarios are less likely to disrupt by tearing, but each has its own challenges, such as no-wall stability limit or minimizing inductive current. Therefore, it is worth developing further AI controllers trained through modified observation, actuation and reward settings to address these different challenges. In addition, the flexibility of the actuators and sensors used in this work at DIII-D will differ from that in ITER and reactors. Control policies under more limited sensing and actuation conditions also need to be developed in the future.

Dynamic model for tearing-instability prediction

To predict tearing events in DIII-D, we first labelled whether each phase was tearing-stable or not (0 or 1) based on the n  = 1 Mirnov coil signal in the experiment. Using this labelled experimental data, we trained a DNN-based multimodal dynamic model that receives various plasma profiles and tokamak actuations as input and predicts the 25-ms-after tearing likelihood as output. The trained dynamic model outputs a continuous value between 0 and 1 (so-called tearability), where a value closer to 1 indicates a higher likelihood of a tearing instability occurring after 25 ms. The architecture of this model is shown in Extended Data Fig. 1 . The detailed descriptions for input and output variables and hyperparameters of the dynamic prediction model can be found in ref. 5 . Although this dynamic model is a black box and cannot explicitly provide the underlying cause of the induced tearing instability, it can be utilized as a surrogate for the response of stability, bypassing expensive real-world experiments. As an example, this dynamic model is used as a training environment for the RL of the tearing-avoidance controller in this work. During the RL training process, the dynamic model predicts future β N and tearability from the given plasma conditions and actuator values determined by the AI controller. Then the reward is estimated based on the predicted state using equation ( 1 ) and provided to the controller as feedback.

Figure 4b–d shows the contour plots of the estimated tearability for possible beam powers at the given plasma conditions of our control experiments. The actual beam power controlled by the AI is indicated by the black solid lines. The dashed lines are the contour line of the threshold value set for each discharge, which can roughly represent the stability limit of the beam power at each point. The plot shows that the trained AI controller proactively avoids touching the tearability threshold before the warning of instability.

The sensitivity of the tearability against the diagnostic errors of the electron temperature and density is shown in Extended Data Fig. 2 . The filled areas in Extended Data Fig. 2 represent the range of tearability predictions when increasing and decreasing the electron temperature and density by 10%, respectively, from the measurements in 193280. The uncertainty in tearability due to electron temperature error is estimated to be, on average, 10%, and the uncertainty due to electron density error is about 20%. However, even when considering diagnostic errors, the trend in tearing stability over time can still be observed to remain consistent.

RL training for tearing avoidance

The dynamic model used for predicting future tearing-instability dynamics is integrated with the OpenAI Gym library 55 , which allows it to interact with the controller as a training environment. The tearing-avoidance controller, another DNN model, is trained using the deep deterministic policy gradient 56 method, which is implemented using Keras-RL ( https://keras.io/ ) 57 .

The observation variables consist of 5 different plasma profiles mapped on 33 equally distributed grids of the magnetic flux coordinate: electron density, electron temperature, ion rotation, safety factor and plasma pressure. The safety factor ( q ) can diverge to infinity at the plasma boundary when the plasma is diverted. Therefore, 1/ q has been used for the observation variables to reduce numerical difficulties 42 . The action variables include the total beam power and the triangularity of the plasma boundary, and their controllable ranges were limited to be consistent with the IBS experiment of DIII-D. The AI-controlled plasma boundary shape has been confirmed to be achievable by the poloidal field coil system of ITER, as shown in Extended Data Fig. 3 .

The RL training process of the AI controller is depicted in Extended Data Fig. 4 . At each iteration, the observation variables (five different profiles) are randomly selected from experimental data. From this observation, the AI controller determines the desirable beam power and plasma triangularity. To reduce the possibility of local optimization, action noises based on the Ornstein–Uhlenbeck process are added to the control action during training. Then the dynamic model predicts β N and tearability after 25 ms based on the given plasma profiles and actuator values. The reward is evaluated according to equation ( 1 ) using the predicted states, and then given as feedback for the RL of the AI controller. As the controller and the dynamic model observe plasma profiles, it can reflect the change of tearing stability even when plasma profiles vary due to unpredictable factors such as wall conditions or impurities. In addition, although this paper focuses on IBS conditions where tearing instability is critical, the RL training itself was not restricted to any specific experimental conditions, ensuring its applicability across all conditions. After training, the Keras-based controller model is converted to C using the Keras2C library 58 for the PCS integration.

Previously, a related work 17 employed a simple bang-bang control scheme using only beam power to handle tearability. Although our control performance may seem similar to that work in terms of β N , it is not true if considering other operating conditions. In ITER and future fusion devices, higher normalized fusion gain ( G   ∝   Q ) with stable core instability is critical. This requires a high β N and small q 95 as \(G\propto {\beta }_{{\rm{N}}}/{q}_{95}^{2}\) . At the same time, owing to limited heating capability, high G has to be achieved with weak plasma rotation (or beam torque). Here, high β N , small \({q}_{95}^{2}\) and low torque are all destabilizing conditions of tearing instability, highlighting tearing instability as a substantial bottleneck of ITER.

As shown in Extended Data Fig. 5 , our control achieves a tearing-stable operation of much higher G than the test experiment shown in ref. 17 . This is possible by maintaining higher (or similar) β N with lower q 95 (4 → 3), where tearing instability is more likely to occur. In addition, this is achieved with a much weaker torque, further highlighting the capability of our RL controller in harsher conditions. Therefore, this work shows more ITER-relevant performance, providing a closer and clearer path to the high fusion gain with robust tearing avoidance in future devices.

In addition, the performance of RL control in achieving high fusion can be further highlighted when considering the non-monotonic effect of β N on tearing instability. Unlike q 95 or torque, both increasing and decreasing β N can destabilize tearing instabilities. This leads to the existence of optimal fusion gain (as G   ∝   β N ), which enables the tearing-stable operation and makes system control more complicated. Here, Extended Data Fig. 6 shows the trace of RL-controller discharge in the space of fusion gain versus time, where the contour colour illustrates the tearability. This clearly shows that the RL controller successfully drives plasma through the valley of tearability, ensuring stable operation and showing its remarkable performance in such a complicated system.

Such a superior performance is feasible by the advantages of RL over conventional approaches, which are described below.

By employing a ‘multi-actuator (beam and shape) multi-objectives (low tearability and high β N )’ controller using RL, we were able to enter a higher -β N region while maintaining tolerable tearability. As shown in Extended Data Fig. 5 , our controlled discharge (193280) shows a higher β N and G than the one in the previous work (176757). This advantage of our controller is because it adjusts the beam and plasma shape simultaneously to achieve both increasing β N and lowering tearability. It is notable that our discharge has more unfavourable conditions (lower q 95 and lower torque) in terms of both β N and tearing stability.

The previous tearability model evaluates the tearing likelihood based on current zero-dimensional measurements, not considering the upcoming actuation control. However, our model considers the one-dimensional detailed profiles and also the upcoming actuations, then predicts the future tearability response to the future control. This can provide a more flexible applicability in terms of control. Our RL controller has been trained to understand this tearability response and can consider future effects, while the previous controller only sees the current stability. By considering the future responses, ours offers a more optimal actuation in the longer term instead of a greedy manner.

This enables the application in more generic situations beyond our experiments. For instance, as shown in Extended Data Fig. 7a , tearability is a nonlinear function of β N . In some cases (Extended Data Fig. 7b ), this relation is also non-monotonic, making increasing the beam power the desired command to reduce tearability (as shown in Extended Data Fig. 7b with a right-directed arrow). This is due to the diversity of the tearing-instability sources such as β N limit, Δ ′ and the current well. In such cases, using a simple control shown in ref. 17 could result in oscillatory actuation or even further destabilization. In the case of RL control, there is less oscillation and it controls more swiftly below the threshold, achieving a higher β N through multi-actuator control, as shown in Extended Data Fig. 7c .

Control of plasma triangularity

Plasma shape parameters are key control knobs that influence various types of plasma instability. In DIII-D, the shape parameters such as triangularity and elongation can be manipulated through proximity control 41 . In this study, we used the top triangularity as one of the action variables for the AI controller. The bottom triangularity remained fixed across our experiments because it is directly linked to the strike point on the inner wall.

We also note that the changes in top triangularity through AI control are quite large compared with typical adjustments. Therefore, it is necessary to verify whether such large plasma shape changes are permitted for the capability of magnetic coils in ITER. Additional analysis, as shown in Extended Data Fig. 3 , confirms that the rescaled plasma shape for ITER can be achieved within the coil current limits.

Robustness of maintaining tearability against different conditions

The experiments in Figs. 3b and 4a have shown that the tearability can be maintained through appropriate AI-based control. However, it is necessary to verify whether it can robustly maintain low tearability when additional actuators are added and plasma conditions change. In particular, ITER plans to use not only 50 MW beams but also 10–20 MW radiofrequency actuators. Electron cyclotron radiofrequency heating directly changes the electron temperature profile and the stability can vary sensitively. Therefore, we conducted an experiment to see whether the AI controller successfully maintains low tearability under new conditions where radiofrequency heating is added. In discharge 193282 (green lines in Extended Data Fig. 8 ), 1.8 MW of radiofrequency heating is preprogrammed to be steadily applied in the background while beam power and plasma triangularity are controlled via AI. Here, the radiofrequency heating is towards the core of the plasma and the current drive at the tearing location is negligible.

However, owing to the sudden loss of plasma current control at t  = 3.1 s, q 95 increased from 3 to 4, and the subsequent discharge did not proceed under the ITER baseline condition. It should be noted that this change in plasma current control was unintentional and not directly related to AI control. Such plasma current fluctuation sharply raised the tearability to exceed the threshold temporarily at t  = 3.2 s, but it was immediately stabilized by continued AI control. Although it is eventually disrupted owing to insufficient plasma current by the loss of plasma current before the preprogrammed end of the flat top, this accidental experiment demonstrates the robustness of AI-based tearability control against additional heating actuators, a wider q 95 range and accidental current fluctuation.

In normal plasma experiments, control parameters are kept stationary with a feed-forward set-up, so that each discharge is a single data point. However, in our experiments, both plasma and control are varying throughout the discharge. Thus, one discharge consists of multiple control cycles. Therefore, our results are more important than one would expect compared with standard fixed control plasma experiments, supporting the reliability of the control scheme.

In addition, the predicted plasma response due to RL control for 1,000 samples randomly selected from the experimental database, which includes not just the IBS but all experimental conditions, is shown in Extended Data Fig. 9a,b . When T  > 0.5 (unstable, top), the controller tries to decrease T rather than affecting β N , and when T  < 0.5 (stable, bottom), it tries to increase β N . This matches the expected response by the reward shown in equation ( 1 ). In 98.6% of the unstable phase, the controller reduced the tearability, and in 90.7% of the stable phase, the controller increased β N .

Extended Data Fig. 9c shows the achieved time-integrated β N for the discharge sequences of our experiment session. Discharges until 193276 either did not have the RL control applied or had tearing instability occurring before the control started, and discharges after 193277 had the RL control applied. Before RL control, all shots except one (193266: low- β N reference shown in Fig. 3b ) were disrupted, but after RL control was applied, only two (193277 and 193282) were disrupted, which were discussed earlier. The average time-integrated β N also increased after the RL control. In addition, the input feature ranges of the controlled discharges are compared with the training database distribution in Extended Data Fig. 10 , which indicates that our experiments are neither too centred (the model not overfitted to our experimental condition) nor too far out (confirming the availability of our controller on the experiments).

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Mnih, V. et al. Human-level control through deep reinforcement learning. Nature 518 , 529–533 (2015).

Article   ADS   CAS   PubMed   Google Scholar  

Cheng, Y. & Zhang, W. Concise deep reinforcement learning obstacle avoidance for underactuated unmanned marine vessels. Neurocomputing 272 , 63–73 (2018).

Article   Google Scholar  

Degrave, J. et al. Magnetic control of tokamak plasmas through deep reinforcement learning. Nature 602 , 414–419 (2022).

Article   ADS   CAS   PubMed   PubMed Central   Google Scholar  

Seo, J. et al. Development of an operation trajectory design algorithm for control of multiple 0D parameters using deep reinforcement learning in KSTAR. Nucl. Fusion 62 , 086049 (2022).

Article   ADS   Google Scholar  

Seo, J. et al. Multimodal prediction of tearing instabilities in a tokamak. In 2023 International Joint Conference on Neural Networks (IJCNN) 1–8 (IEEE, 2023).

Luxon, J. A design retrospective of the DIII-D tokamak. Nucl. Fusion 42 , 614 (2002).

Article   ADS   CAS   Google Scholar  

Betti, R. A milestone in fusion research is reached. Nat. Rev. Phys. 5 , 6–8 (2023).

Han, H. et al. A sustained high-temperature fusion plasma regime facilitated by fast ions. Nature 609 , 269–275 (2022).

Song, Y. et al. Realization of thousand-second improved confinement plasma with Super I-mode in tokamak EAST. Sci. Adv. 9 , eabq5273 (2023).

Article   ADS   PubMed   PubMed Central   Google Scholar  

Mailloux, J. et al. Overview of jet results for optimising ITER operation. Nucl. Fusion 62 , 042026 (2022).

Gibney, E. Nuclear-fusion reactor smashes energy record. Nature 602 , 371 (2022).

Shimada, M. et al. Progress in the ITER physics basis—chapter 1: overview and summary. Nucl. Fusion 47 , S1 (2007).

Article   CAS   Google Scholar  

Schuller, F. C. Disruptions in tokamaks. Plasma Phys. Control. Fusion 37 , A135 (1995).

Kates-Harbeck, J., Svyatkovskiy, A. & Tang, W. Predicting disruptive instabilities in controlled fusion plasmas through deep learning. Nature 568 , 526–531 (2019).

Vega, J. et al. Disruption prediction with artificial intelligence techniques in tokamak plasmas. Nat. Phys. 18 , 741–750 (2022).

Lehnen, M. et al. Disruptions in ITER and strategies for their control and mitigation. J. Nucl. Mater. 463 , 39–48 (2015).

Fu, Y. et al. Machine learning control for disruption and tearing mode avoidance. Phys. Plasmas 27 , 022501 (2020).

de Vries, P. et al. Survey of disruption causes at jet. Nucl. Fusion 51 , 053018 (2011).

Turco, F. et al. The causes of the disruptive tearing instabilities of the ITER baseline scenario in DIII-D. Nucl. Fusion 58 , 106043 (2018).

La Haye, R. J. Neoclassical tearing modes and their control. Phys. Plasmas 13 , 055501 (2006).

Article   ADS   MathSciNet   Google Scholar  

Gantenbein, G. et al. Complete suppression of neoclassical tearing modes with current drive at the electron-cyclotron-resonance frequency in ASDEX upgrade tokamak. Phys. Rev. Lett. 85 , 1242–1245 (2000).

La Haye, R. J. et al. Control of neoclassical tearing modes in DIII-D. Phys. Plasmas 9 , 2051–2060 (2002).

Volpe, F. A. G. et al. Advanced techniques for neoclassical tearing mode control in DIII-D. Phys. Plasmas 16 , 102502 (2009).

Felici, F. et al. Integrated real-time control of MHD instabilities using multi-beam ECRH/ECCD systems on TCV. Nucl. Fusion 52 , 074001 (2012).

Maraschek, M. Control of neoclassical tearing modes. Nucl. Fusion 52 , 074007 (2012).

Kolemen, E. et al. State-of-the-art neoclassical tearing mode control in DIII-D using real-time steerable electron cyclotron current drive launchers. Nucl. Fusion 54 , 073020 (2014).

Park, M., Na, Y.-S., Seo, J., Kim, M. & Kim, K. Effect of electron cyclotron beam width to neoclassical tearing mode stabilization by minimum seeking control in iter. Nucl. Fusion 58 , 016042 (2017).

Bardóczi, L., Logan, N. C. & Strait, E. J. Neoclassical tearing mode seeding by nonlinear three-wave interactions in tokamaks. Phys. Rev. Lett. 127 , 055002 (2021).

Article   ADS   PubMed   Google Scholar  

Zeng, S., Zhu, P., Izzo, V., Li, H. & Jiang, Z. MHD simulations of cold bubble formation from 2/1 tearing mode during massive gas injection in a tokamak. Nucl. Fusion 62 , 026015 (2022).

Yang, X., Liu, Y., Xu, W., He, Y. & Xia, G. Effect of negative triangularity on tearing mode stability in tokamak plasmas. Nucl. Fusion 63 , 066001 (2023).

Wakatsuki, T., Suzuki, T., Hayashi, N., Oyama, N. & Ide, S. Safety factor profile control with reduced central solenoid flux consumption during plasma current ramp-up phase using a reinforcement learning technique. Nucl. Fusion 59 , 066022 (2019).

Seo, J. et al. Feedforward beta control in the KSTAR tokamak by deep reinforcement learning. Nucl. Fusion 61 , 106010 (2021).

Char, I. et al. Offline model-based reinforcement learning for tokamak control. In 2023 Learning for Dynamics and Control Conference (L4DC) 1357–1372 (PMLR, 2023).

Wakatsuki, T., Yoshida, M., Narita, E., Suzuki, T. & Hayashi, N. Simultaneous control of safety factor profile and normalized beta for JT-60SA using reinforcement learning. Nucl. Fusion 63 , 076017 (2023).

Tracey, B. D. et al. Towards practical reinforcement learning for tokamak magnetic control.  Fusion Eng. Des.   200 , 114161 (2024).

Shousha, R. et al. Improved real-time equilibrium reconstruction with kinetic constraints on DIII-D and NSTX-U. In 64th Annual Meeting of the APS Division of Plasma Physics Vol. 67, PP11.00011 (APS, 2022); https://meetings.aps.org/Meeting/DPP22/Session/PP11.11 .

Shousha, R. et al. Machine learning-based real-time kinetic profile reconstruction in DIII-D. Nucl. Fusion 64 , 026006 (2024).

Jalalvand, A., Abbate, J., Conlin, R., Verdoolaege, G. & Kolemen, E. Real-time and adaptive reservoir computing with application to profile prediction in fusion plasma. IEEE Trans. Neural Netw. Learn. Syst. 33 , 2630–2641 (2022).

Article   MathSciNet   PubMed   Google Scholar  

Ferron, J. et al. Real time equilibrium reconstruction for tokamak discharge control. Nucl. Fusion 38 , 1055 (1998).

Boyer, M., Kaye, S. & Erickson, K. Real-time capable modeling of neutral beam injection on NSTX-U using neural networks. Nucl. Fusion 59 , 056008 (2019).

Barr, J. et al. Development and experimental qualification of novel disruption prevention techniques on DIII-D. Nucl. Fusion 61 , 126019 (2021).

Abbate, J., Conlin, R. & Kolemen, E. Data-driven profile prediction for DIII-D. Nucl. Fusion 61 , 046027 (2021).

Seo, J. Solving real-world optimization tasks using physics-informed neural computing. Sci. Rep. 14 , 202 (2024).

Na, Y.-S. et al. Observation of a new type of self-generated current in magnetized plasmas. Nat. Commun. 13 , 6477 (2022).

Carlstrom, T. N. et al. Design and operation of the multipulse Thomson scattering diagnostic on DIII-D (invited). Rev. Sci. Instrum. 63 , 4901–4906 (1992).

Seraydarian, R. P. & Burrell, K. H. Multichordal charge-exchange recombination spectroscopy on the DIII-D tokamak. Rev. Sci. Instrum. 57 , 2012–2014 (1986).

Margo, M. et al. Current state of DIII-D plasma control system. Fusion Eng. Des. 150 , 111368 (2020).

Escande, D. & Ottaviani, M. Simple and rigorous solution for the nonlinear tearing mode. Phys. Lett. A 323 , 278–284 (2004).

Loizu, J. et al. Direct prediction of nonlinear tearing mode saturation using a variational principle. Phys. Plasmas 27 , 070701 (2020).

Muraglia, M. et al. Generation and amplification of magnetic islands by drift interchange turbulence. Phys. Rev. Lett. 107 , 095003 (2011).

Hornsby, W. A. et al. On seed island generation and the non-linear self-consistent interaction of the tearing mode with electromagnetic gyro-kinetic turbulence. Plasma Phys. Control. Fusion 57 , 054018 (2015).

Agullo, O. et al. Nonlinear dynamics of turbulence driven magnetic islands. I. Theoretical aspects. Phys. Plasmas 24 , 042308 (2017).

Choi, G. J. & Hahm, T. S. Long term vortex flow evolution around a magnetic island in tokamaks. Phys. Rev. Lett. 128 , 225001 (2022).

Article   ADS   MathSciNet   CAS   PubMed   Google Scholar  

Sauter, O. et al. Marginal β-limit for neoclassical tearing modes in JET H-mode discharges. Plasma Phys. Control. Fusion 44 , 1999 (2002).

Brockman, G. et al. OpenAI Gym. Preprint at https://arxiv.org/abs/1606.01540 (2016).

Lillicrap, T. P. et al. Continuous control with deep reinforcement learning. Preprint at https://arxiv.org/abs/1509.02971 (2015).

keras-rl2. GitHub https://github.com/inarikami/keras-rl2 (2019).

Conlin, R., Erickson, K., Abbate, J. & Kolemen, E. Keras2c: a library for converting Keras neural networks to real-time compatible C. Eng. App. Artif. Intell. 100 , 104182 (2021).

Download references

Acknowledgements

This material is based on work supported by the US Department of Energy, Office of Science, Office of Fusion Energy Sciences, using the DIII-D National Fusion Facility, a DOE Office of Science user facility, under awards DE-FC02-04ER54698 and DE-AC02-09CH11466. This work was also supported by the National Research Foundation of Korea (NRF) funded by the Korea government (Ministry of Science and ICT) (RS-2023-00255492). Disclaimer: This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favouring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof.

Author information

Authors and affiliations.

Department of Mechanical and Aerospace Engineering, Princeton University, Princeton, NJ, USA

Jaemin Seo, SangKyeun Kim, Azarakhsh Jalalvand, Rory Conlin, Andrew Rothstein, Josiah Wai, Ricardo Shousha & Egemen Kolemen

Department of Physics, Chung-Ang University, Seoul, South Korea

Princeton Plasma Physics Laboratory, Princeton, NJ, USA

SangKyeun Kim, Rory Conlin, Joseph Abbate, Keith Erickson, Ricardo Shousha & Egemen Kolemen

Department of Astrophysical Sciences, Princeton University, Princeton, NJ, USA

Joseph Abbate

You can also search for this author in PubMed   Google Scholar

Contributions

J.S. is the main author of the paper and contributed to developing the controller model, experiments and analyses. S.K. and A.J. contributed equally to writing the paper, developing the controller, experiments and analyses. R.C. contributed to implementing the controller in DIII-D, experiments and analyses. A.R. contributed to developing the controller model, experiments and analyses. J.A. contributed to the experiments. K.E. contributed to implementing the controller in DIII-D. J.W. contributed to the analyses. R.S. contributed to the experiments. E.K. contributed to the conception of this work, experiments, analyses and writing the paper.

Corresponding author

Correspondence to Egemen Kolemen .

Ethics declarations

Competing interests.

The authors declare no competing interests.

Peer review

Peer review information.

Nature thanks Olivier Agullo, Sehyun Kwak and Hiroshi Yamada for their contribution to the peer review of this work.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Extended data figures and tables

Extended data fig. 1 the dnn architecture of the dynamic model that predicts future tearability..

The inputs of the dynamic model are the 1-dimensional signals of the plasma state and the scalar signals of the proposed actuators. The outputs are the normalized plasma pressure ( β N ) and the tearability metric after 25 ms.

Extended Data Fig. 2 The sensitivity of the tearability against the diagnostic errors in 193280.

a , The evolution of tearability with uncertainty range caused by the electron temperature error of 10 %. b , The evolution of tearability with uncertainty range caused by the electron density error of 10 %.

Extended Data Fig. 3 The ITER-rescaled plasma boundary of discharge 193280 and the required poloidal field coil currents.

a , The poloidal cross-section of the ITER first wall, plasma boundaries, and PF coils. The blue shade is the range of the ITER-rescaled plasma boundary of discharge 193280 and the red line is the ITER reference plasma boundary. b , The maximum coil current required to shape each plasma boundary compared to the coil current limits. The PF coils of ITER can support the new plasma boundary shape determined by AI.

Extended Data Fig. 4 The pipeline of the RL training used in our work.

First, random plasma profiles are selected from experimental data to be fed to both the dynamic model and the AI controller. The AI controller observes the plasma profiles and determines the action. Then, the dynamic model predicts the future β N and tearability. Lastly, the reward is estimated from the predicted state to optimize the AI controller.

Extended Data Fig. 5 Comparison of the discharge using a previous controller (176757) and our controlled one (193280).

Multi-actuator multi-objectives control could achieve higher β N and G under more unfavorable condition. Here, the time domain for 176757 was shifted by + 0.75 s to synchronize the H-mode onset between two shots.

Extended Data Fig. 6 Time trace of the normalized fusion gain for discharge 193280, where contour color illustrates the tearability.

The RL control successfully drives plasma through the valley of tearability.

Extended Data Fig. 7 Non-monotonic dependence of tearability and its effect on control.

a , Non-linear dependence of tearability on β N observed in experiments. b , Non-monotonic dependence of tearability on beam power observed in model predictions. c , Comparison of a simple bang-bang controller (black) and our controller (blue) in a simulative plasma. While the simple controller induces an oscillatory actuation, our controller could achieve swifter stabilization with higher β N . The plasma response without adjusting triangularity from the RL control is also shown with blue dashed lines.

Extended Data Fig. 8 Control experiments under the different plasma conditions by adding RF heating.

In the AI-controlled discharge (193282), the plasma current control is suddenly lost at t  = 3.1 s, but the tearability control is still working after that.

Extended Data Fig. 9 Statistics of the predicted plasma response by RL control in the existing database.

a , The response of tearability by control when the original plasma was unstable (top) and stable (bottom). b , The response of β N by control when the original plasma was unstable (top) and stable (bottom). c , Change of the time-integrated β N after the RL control during our experimental session, where circles represent non-disrupted shots, while crosses indicate disrupted ones. After the RL controller was applied, the average time-integrated β N increased, and the disrupted rate decreased.

Extended Data Fig. 10 Comparison of several input data of our experiments with the training database distribution.

a , Radar chart of the major input features distribution space, for the training data (blue) and our experiments (red). b , Time trace of the distribution of selected actuators. c , PCA analysis of the multi-dimensional input data distribution.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ .

Reprints and permissions

About this article

Cite this article.

Seo, J., Kim, S., Jalalvand, A. et al. Avoiding fusion plasma tearing instability with deep reinforcement learning. Nature 626 , 746–751 (2024). https://doi.org/10.1038/s41586-024-07024-9

Download citation

Received : 12 July 2023

Accepted : 03 January 2024

Published : 21 February 2024

Issue Date : 22 February 2024

DOI : https://doi.org/10.1038/s41586-024-07024-9

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

By submitting a comment you agree to abide by our Terms and Community Guidelines . If you find something abusive or that does not comply with our terms or guidelines please flag it as inappropriate.

Quick links

  • Explore articles by subject
  • Guide to authors
  • Editorial policies

Sign up for the Nature Briefing newsletter — what matters in science, free to your inbox daily.

research paper based on artificial intelligence

U.S. flag

An official website of the United States government

The .gov means it’s official. Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

The site is secure. The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

  • Publications
  • Account settings
  • Advanced Search
  • Journal List
  • Future Healthc J
  • v.8(2); 2021 Jul

Logo of futhealthcj

Artificial intelligence in healthcare: transforming the practice of medicine

Junaid bajwa.

A Microsoft Research, Cambridge, UK

Usman Munir

B Microsoft Research, Cambridge, UK

Aditya Nori

C Microsoft Research, Cambridge, UK

Bryan Williams

D University College London, London, UK and director, NIHR UCLH Biomedical Research Centre, London, UK

Artificial intelligence (AI) is a powerful and disruptive area of computer science, with the potential to fundamentally transform the practice of medicine and the delivery of healthcare. In this review article, we outline recent breakthroughs in the application of AI in healthcare, describe a roadmap to building effective, reliable and safe AI systems, and discuss the possible future direction of AI augmented healthcare systems.

Introduction

Healthcare systems around the world face significant challenges in achieving the ‘quadruple aim’ for healthcare: improve population health, improve the patient's experience of care, enhance caregiver experience and reduce the rising cost of care. 1–3 Ageing populations, growing burden of chronic diseases and rising costs of healthcare globally are challenging governments, payers, regulators and providers to innovate and transform models of healthcare delivery. Moreover, against a backdrop now catalysed by the global pandemic, healthcare systems find themselves challenged to ‘perform’ (deliver effective, high-quality care) and ‘transform’ care at scale by leveraging real-world data driven insights directly into patient care. The pandemic has also highlighted the shortages in healthcare workforce and inequities in the access to care, previously articulated by The King's Fund and the World Health Organization (Box ​ (Box1 1 ). 4,5

Workforce challenges in the next decade

The application of technology and artificial intelligence (AI) in healthcare has the potential to address some of these supply-and-demand challenges. The increasing availability of multi-modal data (genomics, economic, demographic, clinical and phenotypic) coupled with technology innovations in mobile, internet of things (IoT), computing power and data security herald a moment of convergence between healthcare and technology to fundamentally transform models of healthcare delivery through AI-augmented healthcare systems.

In particular, cloud computing is enabling the transition of effective and safe AI systems into mainstream healthcare delivery. Cloud computing is providing the computing capacity for the analysis of considerably large amounts of data, at higher speeds and lower costs compared with historic ‘on premises’ infrastructure of healthcare organisations. Indeed, we observe that many technology providers are increasingly seeking to partner with healthcare organisations to drive AI-driven medical innovation enabled by cloud computing and technology-related transformation (Box ​ (Box2 2 ). 6–8

Quotes from technology leaders

Here, we summarise recent breakthroughs in the application of AI in healthcare, describe a roadmap to building effective AI systems and discuss the possible future direction of AI augmented healthcare systems.

What is artificial intelligence?

Simply put, AI refers to the science and engineering of making intelligent machines, through algorithms or a set of rules, which the machine follows to mimic human cognitive functions, such as learning and problem solving. 9 AI systems have the potential to anticipate problems or deal with issues as they come up and, as such, operate in an intentional, intelligent and adaptive manner. 10 AI's strength is in its ability to learn and recognise patterns and relationships from large multidimensional and multimodal datasets; for example, AI systems could translate a patient's entire medical record into a single number that represents a likely diagnosis. 11,12 Moreover, AI systems are dynamic and autonomous, learning and adapting as more data become available. 13

AI is not one ubiquitous, universal technology, rather, it represents several subfields (such as machine learning and deep learning) that, individually or in combination, add intelligence to applications. Machine learning (ML) refers to the study of algorithms that allow computer programs to automatically improve through experience. 14 ML itself may be categorised as ‘supervised’, ‘unsupervised’ and ‘reinforcement learning’ (RL), and there is ongoing research in various sub-fields including ‘semi-supervised’, ‘self-supervised’ and ‘multi-instance’ ML.

  • Supervised learning leverages labelled data (annotated information); for example, using labelled X-ray images of known tumours to detect tumours in new images. 15
  • ‘Unsupervised learning’ attempts to extract information from data without labels; for example, categorising groups of patients with similar symptoms to identify a common cause. 16
  • In RL, computational agents learn by trial and error, or by expert demonstration. The algorithm learns by developing a strategy to maximise rewards. Of note, major breakthroughs in AI in recent years have been based on RL.
  • Deep learning (DL) is a class of algorithms that learns by using a large, many-layered collection of connected processes and exposing these processors to a vast set of examples. DL has emerged as the predominant method in AI today driving improvements in areas such as image and speech recognition. 17,18

How to build effective and trusted AI-augmented healthcare systems?

Despite more than a decade of significant focus, the use and adoption of AI in clinical practice remains limited, with many AI products for healthcare still at the design and develop stage. 19–22 While there are different ways to build AI systems for healthcare, far too often there are attempts to force square pegs into round holes ie find healthcare problems to apply AI solutions to without due consideration to local context (such as clinical workflows, user needs, trust, safety and ethical implications).

We hold the view that AI amplifies and augments, rather than replaces, human intelligence. Hence, when building AI systems in healthcare, it is key to not replace the important elements of the human interaction in medicine but to focus it, and improve the efficiency and effectiveness of that interaction. Moreover, AI innovations in healthcare will come through an in-depth, human-centred understanding of the complexity of patient journeys and care pathways.

In Fig ​ Fig1, 1 , we describe a problem-driven, human-centred approach, adapted from frameworks by Wiens et al , Care and Sendak to building effective and reliable AI-augmented healthcare systems. 23–25

An external file that holds a picture, illustration, etc.
Object name is futurehealth-8-2-e188fig1.jpg

Multi-step, iterative approach to build effective and reliable AI-augmented systems in healthcare.

Design and develop

The first stage is to design and develop AI solutions for the right problems using a human-centred AI and experimentation approach and engaging appropriate stakeholders, especially the healthcare users themselves.

Stakeholder engagement and co-creation

Build a multidisciplinary team including computer and social scientists, operational and research leadership, and clinical stakeholders (physician, caregivers and patients) and subject experts (eg for biomedical scientists) that would include authorisers, motivators, financiers, conveners, connectors, implementers and champions. 26 A multi-stakeholder team brings the technical, strategic, operational expertise to define problems, goals, success metrics and intermediate milestones.

Human-centred AI

A human-centred AI approach combines an ethnographic understanding of health systems, with AI. Through user-designed research, first understand the key problems (we suggest using a qualitative study design to understand ‘what is the problem’, ‘why is it a problem’, ‘to whom does it matter’, ‘why has it not been addressed before’ and ‘why is it not getting attention’) including the needs, constraints and workflows in healthcare organisations, and the facilitators and barriers to the integration of AI within the clinical context. After defining key problems, the next step is to identify which problems are appropriate for AI to solve, whether there is availability of applicable datasets to build and later evaluate AI. By contextualising algorithms in an existing workflow, AI systems would operate within existing norms and practices to ensure adoption, providing appropriate solutions to existing problems for the end user.

Experimentation

The focus should be on piloting of new stepwise experiments to build AI tools, using tight feedback loops from stakeholders to facilitate rapid experiential learning and incremental changes. 27 The experiments would allow the trying out of new ideas simultaneously, exploring to see which one works, learn what works and what doesn't, and why. 28 Experimentation and feedback will help to elucidate the purpose and intended uses for the AI system: the likely end users and the potential harm and ethical implications of AI system to them (for instance, data privacy, security, equity and safety).

Evaluate and validate

Next, we must iteratively evaluate and validate the predictions made by the AI tool to test how well it is functioning. This is critical, and evaluation is based on three dimensions: statistical validity, clinical utility and economic utility.

  • Statistical validity is understanding the performance of AI on metrics of accuracy, reliability, robustness, stability and calibration. High model performance on retrospective, in silico settings is not sufficient to demonstrate clinical utility or impact.
  • To determine clinical utility, evaluate the algorithm in a real-time environment on a hold-out and temporal validation set (eg longitudinal and external geographic datasets) to demonstrate clinical effectiveness and generalisability. 25
  • Economic utility quantifies the net benefit relative to the cost from the investment in the AI system.

Scale and diffuse

Many AI systems are initially designed to solve a problem at one healthcare system based on the patient population specific to that location and context. Scale up of AI systems requires special attention to deployment modalities, model updates, the regulatory system, variation between systems and reimbursement environment.

Monitor and maintain

Even after an AI system has been deployed clinically, it must be continually monitored and maintained to monitor for risks and adverse events using effective post-market surveillance. Healthcare organisations, regulatory bodies and AI developers should cooperate to collate and analyse the relevant datasets for AI performance, clinical and safety-related risks, and adverse events. 29

What are the current and future use cases of AI in healthcare?

AI can enable healthcare systems to achieve their ‘quadruple aim’ by democratising and standardising a future of connected and AI augmented care, precision diagnostics, precision therapeutics and, ultimately, precision medicine (Table ​ (Table1 1 ). 30 Research in the application of AI healthcare continues to accelerate rapidly, with potential use cases being demonstrated across the healthcare sector (both physical and mental health) including drug discovery, virtual clinical consultation, disease diagnosis, prognosis, medication management and health monitoring.

Widescale adoption and application of artificial intelligence in healthcare

Timings are illustrative to widescale adoption of the proposed innovation taking into account challenges / regulatory environment / use at scale.

We describe a non-exhaustive suite of AI applications in healthcare in the near term, medium term and longer term, for the potential capabilities of AI to augment, automate and transform medicine.

AI today (and in the near future)

Currently, AI systems are not reasoning engines ie cannot reason the same way as human physicians, who can draw upon ‘common sense’ or ‘clinical intuition and experience’. 12 Instead, AI resembles a signal translator, translating patterns from datasets. AI systems today are beginning to be adopted by healthcare organisations to automate time consuming, high volume repetitive tasks. Moreover, there is considerable progress in demonstrating the use of AI in precision diagnostics (eg diabetic retinopathy and radiotherapy planning).

AI in the medium term (the next 5–10 years)

In the medium term, we propose that there will be significant progress in the development of powerful algorithms that are efficient (eg require less data to train), able to use unlabelled data, and can combine disparate structured and unstructured data including imaging, electronic health data, multi-omic, behavioural and pharmacological data. In addition, healthcare organisations and medical practices will evolve from being adopters of AI platforms, to becoming co-innovators with technology partners in the development of novel AI systems for precision therapeutics.

AI in the long term (>10 years)

In the long term, AI systems will become more intelligent , enabling AI healthcare systems achieve a state of precision medicine through AI-augmented healthcare and connected care. Healthcare will shift from the traditional one-size-fits-all form of medicine to a preventative, personalised, data-driven disease management model that achieves improved patient outcomes (improved patient and clinical experiences of care) in a more cost-effective delivery system.

Connected/augmented care

AI could significantly reduce inefficiency in healthcare, improve patient flow and experience, and enhance caregiver experience and patient safety through the care pathway; for example, AI could be applied to the remote monitoring of patients (eg intelligent telehealth through wearables/sensors) to identify and provide timely care of patients at risk of deterioration.

In the long term, we expect that healthcare clinics, hospitals, social care services, patients and caregivers to be all connected to a single, interoperable digital infrastructure using passive sensors in combination with ambient intelligence. 31 Following are two AI applications in connected care.

Virtual assistants and AI chatbots

AI chatbots (such as those used in Babylon ( www.babylonhealth.com ) and Ada ( https://ada.com )) are being used by patients to identify symptoms and recommend further actions in community and primary care settings. AI chatbots can be integrated with wearable devices such as smartwatches to provide insights to both patients and caregivers in improving their behaviour, sleep and general wellness.

Ambient and intelligent care

We also note the emergence of ambient sensing without the need for any peripherals.

  • Emerald ( www.emeraldinno.com ): a wireless, touchless sensor and machine learning platform for remote monitoring of sleep, breathing and behaviour, founded by Massachusetts Institute of Technology faculty and researchers.
  • Google nest: claiming to monitor sleep (including sleep disturbances like cough) using motion and sound sensors. 32
  • A recently published article exploring the ability to use smart speakers to contactlessly monitor heart rhythms. 33
  • Automation and ambient clinical intelligence: AI systems leveraging natural language processing (NLP) technology have the potential to automate administrative tasks such as documenting patient visits in electronic health records, optimising clinical workflow and enabling clinicians to focus more time on caring for patients (eg Nuance Dragon Ambient eXperience ( www.nuance.com/healthcare/ambient-clinical-intelligence.html )).

Precision diagnostics

Diagnostic imaging.

The automated classification of medical images is the leading AI application today. A recent review of AI/ML-based medical devices approved in the USA and Europe from 2015–2020 found that more than half (129 (58%) devices in the USA and 126 (53%) devices in Europe) were approved or CE marked for radiological use. 34 Studies have demonstrated AI's ability to meet or exceed the performance of human experts in image-based diagnoses from several medical specialties including pneumonia in radiology (a convolutional neural network trained with labelled frontal chest X-ray images outperformed radiologists in detecting pneumonia), dermatology (a convolutional neural network was trained with clinical images and was found to classify skin lesions accurately), pathology (one study trained AI algorithms with whole-slide pathology images to detect lymph node metastases of breast cancer and compared the results with those of pathologists) and cardiology (a deep learning algorithm diagnosed heart attack with a performance comparable with that of cardiologists). 35–38

We recognise that there are some exemplars in this area in the NHS (eg University of Leeds Virtual Pathology Project and the National Pathology Imaging Co-operative) and expect widescale adoption and scaleup of AI-based diagnostic imaging in the medium term. 39 We provide two use cases of such technologies.

Diabetic retinopathy screening

Key to reducing preventable, diabetes-related vision loss worldwide is screening individuals for detection and the prompt treatment of diabetic retinopathy. However, screening is costly given the substantial number of diabetes patients and limited manpower for eye care worldwide. 40 Research studies on automated AI algorithms for diabetic retinopathy in the USA, Singapore, Thailand and India have demonstrated robust diagnostic performance and cost effectiveness. 41–44 Moreover, Centers for Medicare & Medicaid Services approved Medicare reimbursement for the use of Food and Drug Administration approved AI algorithm ‘IDx-DR’, which demonstrated 87% sensitivity and 90% specificity for detecting more-than-mild diabetic retinopathy. 45

Improving the precision and reducing waiting timings for radiotherapy planning

An important AI application is to assist clinicians for image preparation and planning tasks for radiotherapy cancer treatment. Currently, segmentation of the images is time consuming and laborious task, performed manually by an oncologist using specially designed software to draw contours around the regions of interest. The AI-based InnerEye open-source technology can cut this preparation time for head and neck, and prostate cancer by up to 90%, meaning that waiting times for starting potentially life-saving radiotherapy treatment can be dramatically reduced (Fig ​ (Fig2 2 ). 46,47

An external file that holds a picture, illustration, etc.
Object name is futurehealth-8-2-e188fig2.jpg

Potential applications for the InnerEye deep learning toolkit include quantitative radiology for monitoring tumour progression, planning for surgery and radiotherapy planning. 47

Precision therapeutics

To make progress towards precision therapeutics, we need to considerably improve our understanding of disease. Researchers globally are exploring the cellular and molecular basis of disease, collecting a range of multimodal datasets that can lead to digital and biological biomarkers for diagnosis, severity and progression. Two important future AI applications include immunomics / synthetic biology and drug discovery.

Immunomics and synthetic biology

Through the application of AI tools on multimodal datasets in the future, we may be able to better understand the cellular basis of disease and the clustering of diseases and patient populations to provide more targeted preventive strategies, for example, using immunomics to diagnose and better predict care and treatment options. This will be revolutionary for multiple standards of care, with particular impact in the cancer, neurological and rare disease space, personalising the experience of care for the individual.

AI-driven drug discovery

AI will drive significant improvement in clinical trial design and optimisation of drug manufacturing processes, and, in general, any combinatorial optimisation process in healthcare could be replaced by AI. We have already seen the beginnings of this with the recent announcements by DeepMind and AlphaFold, which now sets the stage for better understanding disease processes, predicting protein structures and developing more targeted therapeutics (for both rare and more common diseases; Fig ​ Fig3 3 ). 48,49

An external file that holds a picture, illustration, etc.
Object name is futurehealth-8-2-e188fig3.jpg

An overview of the main neural network model architecture for AlphaFold. 49 MSA = multiple sequence alignment.

Precision medicine

New curative therapies.

Over the past decade, synthetic biology has produced developments like CRISPR gene editing and some personalised cancer therapies. However, the life cycle for developing such advanced therapies is still extremely inefficient and expensive.

In future, with better access to data (genomic, proteomic, glycomic, metabolomic and bioinformatic), AI will allow us to handle far more systematic complexity and, in turn, help us transform the way we understand, discover and affect biology. This will improve the efficiency of the drug discovery process by helping better predict early which agents are more likely to be effective and also better anticipate adverse drug effects, which have often thwarted the further development of otherwise effective drugs at a costly late stage in the development process. This, in turn will democratise access to novel advanced therapies at a lower cost.

AI empowered healthcare professionals

In the longer term, healthcare professionals will leverage AI in augmenting the care they provide, allowing them to provide safer, standardised and more effective care at the top of their licence; for example, clinicians could use an ‘AI digital consult’ to examine ‘digital twin’ models of their patients (a truly ‘digital and biomedical’ version of a patient), allowing them to ‘test’ the effectiveness, safety and experience of an intervention (such as a cancer drug) in the digital environment prior to delivering the intervention to the patient in the real world.

We recognise that there are significant challenges related to the wider adoption and deployment of AI into healthcare systems. These challenges include, but are not limited to, data quality and access, technical infrastructure, organisational capacity, and ethical and responsible practices in addition to aspects related to safety and regulation. Some of these issues have been covered, but others go beyond the scope of this current article.

Conclusion and key recommendations

Advances in AI have the potential to transform many aspects of healthcare, enabling a future that is more personalised, precise, predictive and portable. It is unclear if we will see an incremental adoption of new technologies or radical adoption of these technological innovations, but the impact of such technologies and the digital renaissance they bring requires health systems to consider how best they will adapt to the changing landscape. For the NHS, the application of such technologies truly has the potential to release time for care back to healthcare professionals, enabling them to focus on what matters to their patients and, in the future, leveraging a globally democratised set of data assets comprising the ‘highest levels of human knowledge’ to ‘work at the limits of science’ to deliver a common high standard of care, wherever and whenever it is delivered, and by whoever. 50 Globally, AI could become a key tool for improving health equity around the world.

As much as the last 10 years have been about the roll out of digitisation of health records for the purposes of efficiency (and in some healthcare systems, billing/reimbursement), the next 10 years will be about the insight and value society can gain from these digital assets, and how these can be translated into driving better clinical outcomes with the assistance of AI, and the subsequent creation of novel data assets and tools. It is clear that we are at an turning point as it relates to the convergence of the practice of medicine and the application of technology, and although there are multiple opportunities, there are formidable challenges that need to be overcome as it relates to the real world and the scale of implementation of such innovation. A key to delivering this vision will be an expansion of translational research in the field of healthcare applications of artificial intelligence. Alongside this, we need investment into the upskilling of a healthcare workforce and future leaders that are digitally enabled, and to understand and embrace, rather than being intimidated by, the potential of an AI-augmented healthcare system.

Healthcare leaders should consider (as a minimum) these issues when planning to leverage AI for health:

  • processes for ethical and responsible access to data: healthcare data is highly sensitive, inconsistent, siloed and not optimised for the purposes of machine learning development, evaluation, implementation and adoption
  • access to domain expertise / prior knowledge to make sense and create some of the rules which need to be applied to the datasets (to generate the necessary insight)
  • access to sufficient computing power to generate decisions in real time, which is being transformed exponentially with the advent of cloud computing
  • research into implementation: critically, we must consider, explore and research issues which arise when you take the algorithm and put it in the real world, building ‘trusted’ AI algorithms embedded into appropriate workflows.
  • See us on facebook
  • See us on twitter
  • See us on youtube
  • See us on linkedin
  • See us on instagram

Stanford Medicine study identifies distinct brain organization patterns in women and men

Stanford Medicine researchers have developed a powerful new artificial intelligence model that can distinguish between male and female brains.

February 20, 2024

sex differences in brain

'A key motivation for this study is that sex plays a crucial role in human brain development, in aging, and in the manifestation of psychiatric and neurological disorders,' said Vinod Menon. clelia-clelia

A new study by Stanford Medicine investigators unveils a new artificial intelligence model that was more than 90% successful at determining whether scans of brain activity came from a woman or a man.

The findings, published Feb. 20 in the Proceedings of the National Academy of Sciences, help resolve a long-term controversy about whether reliable sex differences exist in the human brain and suggest that understanding these differences may be critical to addressing neuropsychiatric conditions that affect women and men differently.

“A key motivation for this study is that sex plays a crucial role in human brain development, in aging, and in the manifestation of psychiatric and neurological disorders,” said Vinod Menon , PhD, professor of psychiatry and behavioral sciences and director of the Stanford Cognitive and Systems Neuroscience Laboratory . “Identifying consistent and replicable sex differences in the healthy adult brain is a critical step toward a deeper understanding of sex-specific vulnerabilities in psychiatric and neurological disorders.”

Menon is the study’s senior author. The lead authors are senior research scientist Srikanth Ryali , PhD, and academic staff researcher Yuan Zhang , PhD.

“Hotspots” that most helped the model distinguish male brains from female ones include the default mode network, a brain system that helps us process self-referential information, and the striatum and limbic network, which are involved in learning and how we respond to rewards.

The investigators noted that this work does not weigh in on whether sex-related differences arise early in life or may be driven by hormonal differences or the different societal circumstances that men and women may be more likely to encounter.

Uncovering brain differences

The extent to which a person’s sex affects how their brain is organized and operates has long been a point of dispute among scientists. While we know the sex chromosomes we are born with help determine the cocktail of hormones our brains are exposed to — particularly during early development, puberty and aging — researchers have long struggled to connect sex to concrete differences in the human brain. Brain structures tend to look much the same in men and women, and previous research examining how brain regions work together has also largely failed to turn up consistent brain indicators of sex.

test

Vinod Menon

In their current study, Menon and his team took advantage of recent advances in artificial intelligence, as well as access to multiple large datasets, to pursue a more powerful analysis than has previously been employed. First, they created a deep neural network model, which learns to classify brain imaging data: As the researchers showed brain scans to the model and told it that it was looking at a male or female brain, the model started to “notice” what subtle patterns could help it tell the difference.

This model demonstrated superior performance compared with those in previous studies, in part because it used a deep neural network that analyzes dynamic MRI scans. This approach captures the intricate interplay among different brain regions. When the researchers tested the model on around 1,500 brain scans, it could almost always tell if the scan came from a woman or a man.

The model’s success suggests that detectable sex differences do exist in the brain but just haven’t been picked up reliably before. The fact that it worked so well in different datasets, including brain scans from multiple sites in the U.S. and Europe, make the findings especially convincing as it controls for many confounds that can plague studies of this kind.

“This is a very strong piece of evidence that sex is a robust determinant of human brain organization,” Menon said.

Making predictions

Until recently, a model like the one Menon’s team employed would help researchers sort brains into different groups but wouldn’t provide information about how the sorting happened. Today, however, researchers have access to a tool called “explainable AI,” which can sift through vast amounts of data to explain how a model’s decisions are made.

Using explainable AI, Menon and his team identified the brain networks that were most important to the model’s judgment of whether a brain scan came from a man or a woman. They found the model was most often looking to the default mode network, striatum, and the limbic network to make the call.

The team then wondered if they could create another model that could predict how well participants would do on certain cognitive tasks based on functional brain features that differ between women and men. They developed sex-specific models of cognitive abilities: One model effectively predicted cognitive performance in men but not women, and another in women but not men. The findings indicate that functional brain characteristics varying between sexes have significant behavioral implications.

“These models worked really well because we successfully separated brain patterns between sexes,” Menon said. “That tells me that overlooking sex differences in brain organization could lead us to miss key factors underlying neuropsychiatric disorders.”

While the team applied their deep neural network model to questions about sex differences, Menon says the model can be applied to answer questions regarding how just about any aspect of brain connectivity might relate to any kind of cognitive ability or behavior. He and his team plan to make their model publicly available for any researcher to use.

“Our AI models have very broad applicability,” Menon said. “A researcher could use our models to look for brain differences linked to learning impairments or social functioning differences, for instance — aspects we are keen to understand better to aid individuals in adapting to and surmounting these challenges.”

The research was sponsored by the National Institutes of Health (grants MH084164, EB022907, MH121069, K25HD074652 and AG072114), the Transdisciplinary Initiative, the Uytengsu-Hamilton 22q11 Programs, the Stanford Maternal and Child Health Research Institute, and the NARSAD Young Investigator Award.

About Stanford Medicine

Stanford Medicine is an integrated academic health system comprising the Stanford School of Medicine and adult and pediatric health care delivery systems. Together, they harness the full potential of biomedicine through collaborative research, education and clinical care for patients. For more information, please visit med.stanford.edu .

Artificial intelligence

Exploring ways AI is applied to health care

Stanford Medicine Magazine: AI

Search code, repositories, users, issues, pull requests...

Provide feedback.

We read every piece of feedback, and take your input very seriously.

Saved searches

Use saved searches to filter your results more quickly.

To see all available qualifiers, see our documentation .

  • Notifications

AAAI 2024 Papers: Explore a comprehensive collection of innovative research papers presented at one of the premier artificial intelligence conferences. Seamlessly integrate code implementations for better understanding. ⭐ experience the forefront of progress in artificial intelligence with this repository!

DmitryRyumin/AAAI-2024-Papers

Folders and files, repository files navigation.

AAAI 2024 Papers: Explore a comprehensive collection of innovative research papers presented at AAAI 2024 , one of the premier artificial intelligence conferences. Seamlessly integrate code implementations for better understanding. ⭐ experience the forefront of progress in artificial intelligence with this repository!

AAAI 2024

The PDF version of the AAAI 2024 Conference Programme , includes a comprehensive table listing all accepted full papers in the main track.

research paper based on artificial intelligence

Conference table will be up to date all the time.

Contributors

research paper based on artificial intelligence

Contributions to improve the completeness of this list are greatly appreciated. If you come across any overlooked papers, please feel free to create pull requests , open issues or contact me via email . Your participation is crucial to making this repository even better.

Papers will be sorted by category as soon as the proceedings are available.

Will soon be added

Star History

Star History Chart

Code of conduct

  • Python 87.1%
  • Shell 12.9%

Help | Advanced Search

Computer Science > Artificial Intelligence

Title: an interactive agent foundation model.

Abstract: The development of artificial intelligence systems is transitioning from creating static, task-specific models to dynamic, agent-based systems capable of performing well in a wide range of applications. We propose an Interactive Agent Foundation Model that uses a novel multi-task agent training paradigm for training AI agents across a wide range of domains, datasets, and tasks. Our training paradigm unifies diverse pre-training strategies, including visual masked auto-encoders, language modeling, and next-action prediction, enabling a versatile and adaptable AI framework. We demonstrate the performance of our framework across three separate domains -- Robotics, Gaming AI, and Healthcare. Our model demonstrates its ability to generate meaningful and contextually relevant outputs in each area. The strength of our approach lies in its generality, leveraging a variety of data sources such as robotics sequences, gameplay data, large-scale video datasets, and textual information for effective multimodal and multi-task learning. Our approach provides a promising avenue for developing generalist, action-taking, multimodal systems.

Submission history

Access paper:.

  • Download PDF
  • Other Formats

license icon

References & Citations

  • Google Scholar
  • Semantic Scholar

BibTeX formatted citation

BibSonomy logo

Bibliographic and Citation Tools

Code, data and media associated with this article, recommenders and search tools.

  • Institution

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs .

COMMENTS

  1. Artificial intelligence: A powerful paradigm for scientific research

    This paper undertakes a comprehensive survey on the development and application of AI in different aspects of fundamental sciences, including information science, mathematics, medical science, materials science, geoscience, life science, physics, and chemistry.

  2. The impact of artificial intelligence on human society and bioethics

    The new development of the long-term goal of many researchers is to create strong AI or artificial general intelligence (AGI) which is the speculative intelligence of a machine that has the capacity to understand or learn any intelligent task human being can, thus assisting human to unravel the confronted problem.

  3. PDF The Impact of Artificial Intelligence on Innovation

    The Impact of Artificial Intelligence on Innovation Iain M. Cockburn, Rebecca Henderson, and Scott Stern NBER Working Paper No. 24449 March 2018 JEL No. L1 ABSTRACT Artificial intelligence may greatly increase the efficiency of the existing economy.

  4. The present and future of AI

    The 2021 report is the second in a series that will be released every five years until 2116. Titled "Gathering Strength, Gathering Storms," the report explores the various ways AI is increasingly touching people's lives in settings that range from movie recommendations and voice assistants to autonomous driving and automated medical ...

  5. AI technologies for education: Recent research & future directions

    From unique educational perspectives, this article reports a comprehensive review of selected empirical studies on artificial intelligence in education (AIEd) published in 1993-2020, as collected in the Web of Sciences database and selected AIEd-specialized journals.

  6. 578339 PDFs

    Artificial Intelligence | Explore the latest full-text research PDFs, articles, conference papers, preprints and more on ARTIFICIAL INTELLIGENCE. Find methods information, sources,...

  7. Machine Learning: Algorithms, Real-World Applications and Research

    1 Mention Explore all metrics Abstract In the current age of the Fourth Industrial Revolution (4 IR or Industry 4.0), the digital world has a wealth of data, such as Internet of Things (IoT) data, cybersecurity data, mobile data, business data, social media data, health data, etc.

  8. Cognitive psychology-based artificial intelligence review

    Introduction. At present, in the development of artificial intelligence (AI), the scientific community is mostly based on brain cognition research (Nadji-Tehrani and Eslami, 2020), which is to reproduce the real physiological activities of our human brain through computer software.This replication of the biology of the human brain cannot well simulate the subjective psychological changes ...

  9. Artificial intelligence and machine learning research: towards digital

    A variety of innovative topics are included in the agenda of the published papers in this special issue including topics such as: Stock market Prediction using Machine learning. Detection of Apple Diseases and Pests based on Multi-Model LSTM-based Convolutional Neural Networks. ML for Searching. Machine Learning for Learning Automata

  10. AIJ

    The journal of Artificial Intelligence (AIJ) welcomes papers on broad aspects of AI that constitute advances in the overall field including, but not limited to, cognition and AI, automated reasoning and inference, case-based reasoning, commonsense reasoning, computer vision, constraint processing, ethical AI, heuristic search, human interfaces, ...

  11. Frontiers in Artificial Intelligence

    A nexus for research in core and applied AI areas, this journal focuses on the enormous expansion of AI into aspects of modern life such as finance, law, medicine, agriculture, and human learning. ... Artificial Intelligence in Finance; Big Data and AI in High Energy Physics; Fuzzy Systems; ... Advancing Knowledge-Based Economies and Societies ...

  12. The impact of artificial intelligence on scientific practices: an

    1. Introduction. Artificial intelligence (AI) is now a major driver of societal acceleration (Gruetzemacher & Whittlestone, Citation 2022), making a significant impact on science (e.g. AAAS, Citation 2021) and science education (e.g. Alasadi & Baiz, Citation 2023; Cooper, Citation 2023) raising questions, for example about the use of AI tools in teaching (Talanquer, Citation 2023) and ...

  13. Artificial intelligence Internet of Things: A new paradigm of

    The survey work has used relevant papers which were searched from databases and other sources including IEEE Xplore, ScienceDirect, Scopus, and Google Scholar. ... et al. Artificial intelligence-based sensors for next generation IoT applications: a review. IEEE Sens J 2021; 21: 24920 ... et al. Research on artificial intelligence enhancing ...

  14. Artificial Intelligence in Education (AIEd): a high-level academic and

    In the past few decades, technology has completely transformed the world around us. Indeed, experts believe that the next big digital transformation in how we live, communicate, work, trade and learn will be driven by Artificial Intelligence (AI) [83]. This paper presents a high-level industrial and academic overview of AI in Education (AIEd). It presents the focus of latest research in AIEd ...

  15. Artificial Intelligence in the 21st Century

    The field of artificial intelligence (AI) has shown an upward trend of growth in the 21st century (from 2000 to 2015). The evolution in AI has advanced the development of human society in our own time, with dramatic revolutions shaped by both theories and techniques. However, the multidisciplinary and fast-growing features make AI a field in which it is difficult to be well understood. In this ...

  16. Spontaneous Theory of Mind for Artificial Intelligence

    Existing approaches to Theory of Mind (ToM) in Artificial Intelligence (AI) overemphasize prompted, or cue-based, ToM, which may limit our collective ability to develop Artificial Social Intelligence (ASI). Drawing from research in computer science, cognitive science, and related disciplines, we contrast prompted ToM with what we call spontaneous ToM -- reasoning about others' mental states ...

  17. IEEE Transactions on Artificial Intelligence

    IEEE Transactions on Artificial Intelligence. null | IEEE Xplore. Need Help? US & Canada: +1 800 678 4333 Worldwide: +1 732 981 0060 Contact & Support

  18. Journal of Artificial Intelligence Research

    The Journal of Artificial Intelligence Research (JAIR) is dedicated to the rapid dissemination of important research results to the global artificial intelligence (AI) community. The journal's scope encompasses all areas of AI, including agents and multi-agent systems, automated reasoning, constraint processing and search, knowledge ...

  19. Employing nano-enabled artificial intelligence (AI)-based smart

    Cancer has been classified as a diverse illness with a wide range of subgroups. Its early identification and prognosis, which have become a requirement of cancer research, are essential for clinical treatment. Patients have already benefited greatly from the use of artificial intelligence (AI), machine learning (ML Recent Review Articles Nanoscale 2024 Emerging Investigators

  20. Avoiding fusion plasma tearing instability with deep ...

    This presents an obstacle-avoidance problem for which artificial intelligence based on reinforcement learning has recently shown remarkable performance 1,2,3,4. However, the obstacle here, the ...

  21. Artificial intelligence in healthcare: transforming the practice of

    Artificial intelligence (AI) is a powerful and disruptive area of computer science, with the potential to fundamentally transform the practice of medicine and the delivery of healthcare.

  22. (PDF) Research paper on Artificial Intelligence

    Research paper on Artificial Intelligence December 2022 Authors: Ashutosh Kumar Galgotias University Rachna Priya Swarna Kumari Discover the world's research 25+ million members 160+...

  23. Artificial intelligence in information systems research: A systematic

    In doing so, this study makes important contributions, namely (i) an identification of the current reported business value and contributions of AI, (ii) research and practical implications on the use of AI and (iii) opportunities for future AI research in the form of a research agenda.

  24. Stanford Medicine study identifies distinct brain organization patterns

    A new study by Stanford Medicine investigators unveils a new artificial intelligence model that was more than 90% successful at determining whether scans of brain activity came from a woman or a man.. The findings, published Feb. 20 in the Proceedings of the National Academy of Sciences, help resolve a long-term controversy about whether reliable sex differences exist in the human brain and ...

  25. Artificial Intelligence as a Tool Supporting Prayer Practices

    At the same time, the conducted research indicated that artificial intelligence, rated on a scale of 1-5, could become an object of cult similar to religious worship (2.74); it could replace religious practices (2.39); respondents believed more in the capabilities of AI than in the abilities of God (2.30); and respondents considered AI as ...

  26. DmitryRyumin/AAAI-2024-Papers

    AAAI 2024 Papers: Explore a comprehensive collection of innovative research papers presented at one of the premier artificial intelligence conferences. Seamlessly integrate code implementations for better understanding. ⭐ experience the forefront of progress in artificial intelligence with this repository!

  27. [2402.05929] An Interactive Agent Foundation Model

    The development of artificial intelligence systems is transitioning from creating static, task-specific models to dynamic, agent-based systems capable of performing well in a wide range of applications. We propose an Interactive Agent Foundation Model that uses a novel multi-task agent training paradigm for training AI agents across a wide range of domains, datasets, and tasks. Our training ...

  28. (PDF) Research Paper on Artificial Intelligence

    June 2013. Nimish Kumar. Artificial Intelligence (AI) is the branch of computer science concerned with building machines (computers) that simulate human behavior. The term was coined in 1956 by ...

  29. Artificial intelligence: A powerful paradigm for scientific research

    Artificial intelligence: A powerful paradigm for scientific research - ScienceDirect Volume 2, Issue 4, 28 November 2021, 100179 Review Artificial intelligence: A powerful paradigm for scientific research Yongjun Xu 1 35 42 , Xin Liu 5 35 42 , Xin Cao 10 42 , Changping Huang 18 35 42 , Enke Liu 11 37 42 , Sen Qian 26 42 , Xingchen Liu 28 42 ,

  30. Improving the quality of big data using machine learning and artificial

    Currently, due to the large and continuous increase in the volume of data, there are systems and applications developed to deal with the preparation of large data in storage, management and processing, and the traditional methods are not suitable for dealing with it due to the large and complex size Which leads to consumption of time and lack of efficiency and performance of the techniques ...