《美国白宫：2024人工智能技术对全球研究的潜在影响研究报告-加速研究：利用AI应对全球挑战（英文版）（63页）.pdf》由会员分享，可在线阅读，更多相关《美国白宫：2024人工智能技术对全球研究的潜在影响研究报告-加速研究：利用AI应对全球挑战（英文版）（63页）.pdf（63页珍藏版）》请在三个皮匠报告上搜索。

1、 REPORT TO THE PRESIDENT Supercharging Research:Harnessing Artificial Intelligence to Meet Global Challenges Executive Office of the President Presidents Council of Advisors on Science and Technology APRIL 2024 1 About the Presidents Council of Advisors on Science and Technology The Presidents Counc

2、il of Advisors on Science and Technology(PCAST)is a federal advisory committee appointed by the President to augment the science and technology advice available to him from inside the White House and from the federal agencies.PCAST comprises of 28 of the Nations thought leaders,selected for their di

3、stinguished service and accomplishments in academia,government,and the private sector.PCAST advises the President on matters involving science,technology,and innovation policy,as well as on matters involving scientific and technological information that is needed to inform policy affecting the econo

4、my,worker empowerment,education,energy,the environment,public health,national and homeland security,racial equity,and other topics.For more information about PCAST see www.whitehouse.gov/pcast.2 EXECUTIVE OFFICE OF THE PRESIDENT PRESIDENTS COUNCIL OF ADVISORS ON SCIENCE AND TECHNOLOGY WASHINGTON,D.C

5、.20502 President Joseph R.Biden,Jr.The White House Washington,D.C.Dear Mr.President,Your Presidents Council of Advisors on Science and Technology(PCAST)is excited by the forward-looking approach that your Administration has taken to advance the safe and effective use of artificial intelligence(AI).1

6、,2,3 As requested in your landmark Executive Order on the Safe,Secure,and Trustworthy Development and Use of Artificial Intelligence,we are pleased to report here on the possibilities that AI can enable when applied to research to address major societal and global challenges.AI will fundamentally tr

7、ansform the way we do science.Researchers in many fields are already employing AI to identify new solutions to a wide array of long-standing problems.Today,scientists and engineers are using AI to envision,predictively design,and create novel materials and therapeutic drugs.In the near future,AI wil

8、l enable unprecedented advances in the social sciences,both through new methods of analyzing existing data and the development and analysis of new kinds of anonymized and validated data.Such advances will allow government to better understand how policies affect the American people,and improve those

9、 policies to better meet societal needs and challenges.AI will also allow researchers to run millions of computer-based simulated experiments quickly to provide guidance about the most important real-world experiments to run.In industrial laboratories,rich simulations will be able to identify hazard

10、s or faults in design so that scientists and engineers can create safer,scalable,and efficient products that American industry and American consumers can depend on.In sum,AI is revolutionizing the research process,enriching scientific models,and accelerating data generation and analysis,with impacts

11、 that will be far-ranging.In addition to its opportunities,we must recognize that AI can create new issues and challenges,such as distilling errors and biases embedded in skewed training data,the enormousand increasingamounts of energy required for the computational processes,the possibility that fa

12、ulty science could be unwittingly generated,and the ease with which nefarious actors could use new powerful AI technologies for malicious purposes.Expert human supervision,building protections into the AI algorithms,and a culture of responsible use that includes appropriate application of regulatory

13、 frameworks,as outlined in your Blueprint for an AI Bill of Rights and the AI Risk Management Framework from the National Institute of Standards and Technology,will be essential to mitigating the weaknesses and dangers of AI.Fortunately,reproducibility and validation are core tenets of the 1 The Whi

14、te House.(2022 October).Blueprint for an AI Bill of Rights:Making Automated Systems Work for The American People.2 Executive Order 14110,88 FR 75191.(2023 October).Safe,Secure,and Trustworthy Development and Use of Artificial Intelligence.3 The White House Office of Management and Budget.(2024 March

15、).Advancing Governance,Innovation,and Risk Management for Agency Use of Artificial Intelligence.Memo M-24-10 3 scientific method.As such,the scientific community is already engaged in vibrant and pathbreaking research on AI reproducibility,testing,and evaluation so that researchersand everyonewill e

16、ventually be able to use AI tools with the same confidence with which we use calculators.Ultimately,our vision is that the responsible use of AI will empower scientists and engineers to leverage the opportunities of this transformative technology while navigating and mitigating its weaknesses.Per yo

17、ur Executive Order,and building on the bold work of your Administration,PCAST offers findings and recommendations for action that will help the United States to harness the full potential of AI to equitably and responsibly supercharge research to meet the worlds challenges.Sincerely,The Presidents C

18、ouncil of Advisors on Science and Technology 4 Table of Contents Letter to The White House.2 Working Group on Artificial Intelligence.9 Executive Summary.10 1.Introduction.15 An acceleration of science.16 Artificial intelligence and machine learning.17 Generative AI.18 2.A Vision for the Future of A

19、I-Enabled R&D.20 AI methods will help researchers prioritize the most likely solutions.21 By handling routine tasks,AI will allow scientists to focus on core research.21 Rote laboratory processes will be automated and improved.22 Previously intractable simulations will become possible.22 Shared mode

20、ls and data will reduce duplication of effort,democratize research,and reduce the total cost of using AI.23 Multimodal foundation models will bring together multiple forms of data and create new synergies among branches of science.24 AI will help researchers do more with data.25 New forms of collabo

21、ration will emerge.26 Responsible AI practices will be integrated into research workflows.26 Once the necessary AI infrastructure is in place,new scientific moonshots will become possible.27 3.Key Opportunities for AI to Supercharge Discovery and Address Global and Societal Challenges.28 3.1.A Phase

22、 Change for Materials Discovery.28 3.2 AI for Designing Advanced Semiconductors.31 3.3.Understanding and Addressing Climate Change and Extreme Weather.33 3.4 Revealing the Fundamental Physics of the Universe.35 3.5 Studying Human Behavior,Organizations,and Institutions.37 3.6 Advancing Fundamental U

23、nderstandings in the Life Sciences.39 3.7 Breakthrough Applications of AI in the Life Sciences.41 4.Findings&Recommendations.46 Finding 1:Important research is being hindered by lack of access to advanced models.47 5 Recommendation 1:Expand existing efforts to broadly and equitably share basic AI re

24、sources.48 Finding 2:Cutting-edge research requires access to high quality data.50 Recommendation 2:Expand secure access to federal data sets for approved critical research needs,with appropriate protections and safeguards.51 Finding 3:AI provides a unique resource for collaborations across academia

25、,industry,and all branches of the federal government.52 Recommendation 3:Support both basic and applied research in AI that involves collaborations across academia,industry,national and federal laboratories,and federal agencies as outlined in the vision for the NAIRR developed by the NAIRR Task Forc

26、e.53 Finding 4:Without proper benchmark metrics,validation procedures,and responsible practices,AI systems can give unreliable outputs whose quality is difficult to evaluate,and which could be harmful for a scientific field and its applications.54 Recommendation 4:Adopt principles of responsible,tra

27、nsparent,and trustworthy AI use throughout all stages of the scientific research process.54 Finding 5:Optimal performance requires both AI and human expertise.55 Recommendation 5:Encourage innovative approaches to integrating AI assistance into scientific workflows.56 5.Conclusion.57 Appendix A:Glos

28、sary.58 Appendix B:External Experts Consulted.61 Acknowledgments.62 6 The Presidents Council of Advisors on Science and Technology Co-Chairs Frances H.Arnold Linus Pauling Professor of Chemical Engineering,Bioengineering,and Biochemistry California Institute of Technology Arati Prabhakar Director,Of

29、fice of Science and Technology Policy Assistant to the President for Science and Technology The White House Maria T.Zuber Vice President for Research and E.A.Griswold Professor of Geophysics Massachusetts Institute of Technology Members Dan E.Arvizu Former Chancellor New Mexico State University Syst

30、em Dennis Assanis President University of Delaware John Banovetz Executive Vice President,Chief Technology Officer and Environmental Responsibility 3M Company Frances Coln Senior Director,International Climate Center for American Progress Lisa A.Cooper Bloomberg Distinguished Professor of Equity in

31、Health and Healthcare and Director of the Center for Health Equity Johns Hopkins University John O.Dabiri Centennial Professor of Aeronautics and Mechanical Engineering California Institute of Technology William Dally Chief Scientist and Senior Vice President for Research NVIDIA Sue Desmond-Hellmann

32、 Former CEO Bill&Melinda Gates Foundation Inez Fung Professor of Atmospheric Science University of California,Berkeley Andrea Goldsmith Dean of the School of Engineering and Applied Science and the Arthur LeGrand Doty Professor of Electrical and Computer Engineering Princeton University Laura H.Gree

33、ne Chief Scientist,National High Magnetic Field Laboratory Florida State University,University of Florida,Los Alamos National Laboratory Marie Krafft Professor of Physics Florida State University 7 Paula Hammond Institute Professor,Vice Provost for Faculty,and member of the Koch Institute for Integr

34、ative Cancer Research Massachusetts Institute of Technology Eric Horvitz Chief Scientific Officer Microsoft Joe Kiani Chairman and CEO Masimo Jon Levin Philip H.Knight Professor and Dean of the Graduate School of Business Stanford University Steve Pacala Frederick D.Petrie Professor Emeritus in the

35、Department of Ecology and Evolutionary Biology Princeton University Saul Perlmutter Franklin W.and Karen Weber Dabby Professor of Physics and Director of the Berkeley Institute for Data Science University of California,Berkeley Senior Scientist Lawrence Berkeley National Labs William Press Leslie Su

36、rginer Professor of Computer Science and Integrative Biology The University of Texas at Austin Jennifer Richeson Philip R.Allen Professor of Psychology and Director of the Social Perception and Communication Lab Yale University Vicki Sato Professor of Management Practice(Retired)Harvard Business Sch

37、ool Lisa Su President and CEO Advanced Micro Devices(AMD)Kathryn D.Sullivan Former Astronaut National Aeronautics and Space Administration Former Administrator National Oceanic and Atmospheric Administration Terence Tao Professor&the James and Carol Collins Chair in the College of Letters and Scienc

38、es University of California,Los Angeles Phil Venables Chief Information Security Officer Google Cloud Catherine Woteki Visiting Distinguished Institute Professor in the Biocomplexity Institute University of Virginia Professor of Food Science and Human Nutrition Iowa State University 8 PCAST Staff La

39、ra Campbell Executive Director Reba Bandyopadhyay Deputy Executive Director Melissa Edwards Assistant Deputy Executive Director Bich-Thuy(Twee)Sim Assistant Director for Transformative Medicine and Health Innovation Kimberly Lawrence Administrative Specialist Riya Dhar Intern Maya Millette Former In

40、tern 9 Working Group on Artificial Intelligence Working Group members participated in the preparation of this report.The full membership of PCAST reviewed and approved the report.Co-Leads Terence Tao Professor&the James and Carol Collins Chair in the College of Letters and Sciences University of Cal

41、ifornia,Los Angeles Laura H.Greene Chief Scientist,National High Magnetic Field Laboratory Florida State University,University of Florida,Los Alamos National Laboratory Marie Krafft Professor of Physics Florida State University Members John Banovetz Executive Vice President,Chief Technology Officer

42、and Environmental Responsibility 3M Company William Dally Chief Scientist and Senior Vice President for Research NVIDIA Eric Horvitz Chief Scientific Officer Microsoft Jon Levin Philip H.Knight Professor and Dean of the Graduate School of Business Stanford University Saul Perlmutter Franklin W.and K

43、aren Weber Dabby Professor of Physics and Director of the Berkeley Institute for Data Science University of California,Berkeley Senior Scientist Lawrence Berkeley National Labs William Press Leslie Surginer Professor of Computer Science and Integrative Biology The University of Texas at Austin Lisa

44、Su President and CEO Advanced Micro Devices(AMD)Phil Venables Chief Information Security Officer Google Cloud 10 Executive Summary Artificial Intelligence(AI)has the potential to revolutionize our ability to address humanitys most urgent challenges by providing researchers with tools that will accel

45、erate scientific discoveries and technological advances.Generative AI,which can create content based on vast data sets and extensive computation,is poised to be particularly transformative.Examples of generative AI include large language models,image generating models,and generative scientific model

46、s.In his comprehensive Executive Order on the Safe,Secure,and Trustworthy Development and Use of Artificial Intelligence,issued on October 30,2023,President Biden charged PCAST to report on“the potential role of AIin research aimed at tackling major societal and global challenges.”PCAST is pleased t

47、o offer this report in fulfillment of this charge.With well-designed,equitably shared,and responsibly used infrastructure,AI will enable scientists to address urgent challenges,including improving human health and enhancing weather prediction in a time of climate change.AI can help explore long-stan

48、ding scientific mysteries that inspire and stretch human creativity,such as uncovering the origin and evolution of the universe.AI will also help researchers address continuing national needs,from accelerated semiconductor chip design to the discovery of new materials to address our energy needs.Fur

49、thermore,AI is starting to remove barriers that make scientific research slow and expensive,for instance by providing the means for rapidly identifying the best drug candidates for testing(thus reducing the number of expensive laboratory trials),helping to optimize experimental designs,and uncoverin

50、g connections in data much more efficiently than can be done by hand or using traditional data science methods.If basic AI resources,validated data,and scientific tools and training are made broadly accessible,AI technologies have the potential to democratize scientific knowledge,4 bringing intercon

51、nected technical concepts to many more people and enabling diverse researchers to bring their expertise and perspectives to societal and global challenges.Just as with any other new tool or technology,realizing the potentials of AI will require addressing its limitations.These issues include mislead

52、ing or incorrect results,perpetuation of bias or inequity5 and sampling errors from patterns embedded in the model-training data,limited access to high quality training data,the challenges of protecting intellectual property and privacy,the significant energy required to train or deploy a model or r

53、un the AI algorithms,and the risk that bad or nefarious actors will use readily available AI tools for malicious purposes.Many public and private sector activities addressing these issues are already underway,including government efforts tasked under the October 2023 Executive Order on AI.Reproducib

54、ility and validation are key principles underlying scientific integrity and the scientific method and must continue to be held at high value as we develop a culture of responsible AI use and expert human supervision of AI applications.AI has the potential to transform every scientific discipline and

55、 many aspects of the way we conduct science.Scientists are already employing AI to create new functional materials that we presently do not know how to design;these include superconductors and thermoelectric materials which would not only enhance our energy efficiency but also reduce our carbon foot

56、print.In a similar vein,AI models are helping researchers create new designs for manufacturing processes and products,and 4 Dessimoz,C.and Thomas,P.(2024 March).AI and the democratization of knowledge.Scientific Data.5 Birhane,A.(2022 October).The unseen Black faces of AI algorithms.Nature.11 develo

57、p new drug therapies which in the future could enable individualized treatment of specific cancers and viruses.AI models are also helping engineers design semiconductor chips,producing better designs with less human effort and time.In health care,AI technologies are creating new ways to analyze a br

58、oad spectrum of medical data for applications like the early diagnosis of diseases6 that can lead to timely intervention and the detection of medical errors.7 PCAST also foresees widely available AI-powered ultra-personalized medicine tailored to a specific individual and disease process that will i

59、nclude details of medical history,genetic information,and signals,such as how healthy and unhealthy cells are behaving.AI is also transforming science by improving our scientific models.In climate science,AI models are starting to enhance weather prediction,as well as advancing whole-earth models fo

60、r water management,greenhouse gas monitoring,and predicting the impacts of catastrophes.Scientists have already used AI to successfully predict the structure of proteins;new foundation models will unlock more secrets of cellular biology and power computer simulations of intracellular interactions th

61、at can be used to explore new therapies.AI models promise to help us understand the origin of our universe by allowing us to test numerous cosmological hypotheses via rapid simulations.Such AI-enabled modeling may even help scientists discover new laws of physics.AI will enable unprecedented advance

62、s in the social sciences,complementing qualitative methods with new quantitative techniques for analyzing existing data,as well as the development and analysis of newer types of data,e.g.,step counts on smartphones,anonymized data drawn with permission from search and browsing,or images posted on so

63、cial media.AI could supercharge research using vast data sets,such as those that have long been collected and curated by federal statistical agenciesideally complemented by those held by the private sectoras input for designing effective federal policies.Application of AI to these long-standing and

64、newer social science data sets could facilitate more effective,responsive,and fairer data-driven policymaking and delivery of services.These few examples of AI-assisted research illustrate that with the responsible use of AI technology,human scientists will be empowered to realize transformational d

65、iscoveries.Furthermore,PCAST expects that responsible sharing of basic AI resources will help to democratize science and tackle major societal and global challenges.The use of AI for science and technology research is accelerating rapidly across the globe and therefore demands our commitment to U.S.

66、leadership in the applications of this powerful new tool.Building on the work of the Biden Administration,the United States must act boldly and thoughtfully to maintain our nations lead in research,in the innovative applications of AI,and in establishing frameworks and norms for the safe and respons

67、ible use of AI.In this report,PCAST offers five specific findings and recommendations for action that will help the U.S.to harness the full potential of AI to equitably and responsibly supercharge scientific discovery.6 Bera,K.et al.(2019 August).Artificial intelligence in digital pathology-new tool

68、s for diagnosis and precision oncology.Nature Reviews Clinical Oncology.7 Nguyen,V.et al.(2023 November).Efficient automated error detection in medical data using deep-learning and label-clustering.Scientific Reports.12 Summary of Recommendations Recommendation 1:Expand existing efforts to broadly a

69、nd equitably share fundamental AI resources.Extensive support for widely accessible shared models,data sets,benchmarks,and computational resources is essential to ensuring that academic researchers,national and federal laboratories,and smaller companies and non-profit organizations can use AI to cre

70、ate benefits for the nation.In the U.S.,the most promising effort in this direction is the National Artificial Intelligence Research Resource(NAIRR),which is currently a pilot project.PCAST recommends that the NAIRR pilot be expeditiously expanded to the scale envisioned by the NAIRR Task Force8 and

71、 fully funded.The full-scale NAIRR,together with industry partnerships and other AI infrastructure efforts at both the federal and state levels,could serve as a stepping stone towards AI infrastructure projects at the national or international level to facilitate high-impact research.Recommendation

72、2:Expand secure access to federal data sets for approved critical research needs,with appropriate protections and safeguards.The benefits of allowing limited,secure access to federal data sets by approved researchers,as well as allowing the release of carefully anonymized versions of such data sets

73、to curated resource centers such as NAIRR,are immense.PCAST strongly encourages expansion of existing pilot programs for secure data access and the development of guidelines for federal database management that incorporate cutting-edge privacy protection technologies as they become available.There i

74、s great potential to use modern AI technologies to automate aspects of the curation of such data sets.PCAST encourages the use of AI to improve data curation as a long-term goal of federal data sharing initiatives such as data.gov.PCAST endorses the efforts of federal agencies to mandate responsible

75、 sharing of data sets arising from the research that they fund or conduct.9 We encourage further enforcement of such mandates,to include sharing of AI models trained on data from federally funded research,in conjunction with sufficient resources to support the required actions.Recommendation 3:Suppo

76、rt both basic and applied research in AI that involves collaborations across academia,industry,national and federal laboratories,and federal agencies as outlined in the vision for the NAIRR developed by the NAIRR Task Force.The boundaries between federally funded academic research and private sector

77、 research are hazy.Many researchers move among affiliations with academic institutions,non-profit organizations,and/or private companies,and a significant share of all AI research and development(R&D)is 8 National Artificial Intelligence Research Resource Task Force.(2023 January).Strengthening and

78、Democratizing the U.S.Artificial Intelligence Innovation Ecosystem:An Implementation Plan for a National Artificial Intelligence Research Resource.9 The White House Office of Science and Technology Policy.(2022 August).Ensuring Free,Immediate,and Equitable Access to Federally Funded Research.13 curr

79、ently supported by private companies.10 To capitalize fully on the potential benefits of AI for science,research that involves a breadth of promising and productive hypotheses and approaches must be supported.This may require that funding agencies broaden their postures regarding how to work with in

80、dustry and which researchers can be supported in order to facilitate innovative research efforts and collaborations among different sectors.Examples of such collaboration could include creation of high quality curated public scientific data sets from multiple sources or the creation of multimodal fo

81、undation models.11 Recommendation 4:Adopt principles of responsible,transparent,and trustworthy AI use throughout all stages of the scientific research process.Managing the risks of inaccurate,biased,harmful,or non-replicable findings from scientific uses of AI should be planned from the initial sta

82、ges of a research project rather than performed as an afterthought.PCAST recommends that federal funding agencies consider updating their responsible conduct of research guidelines to require plans for responsible AI use from researchers.These plans should include recommended best practices from age

83、ncy offices and committees that address potential AI-related risks and describe the supervision procedures for use of any automated process.12 To minimize additional administrative burden on researchers and build a culture of responsibility,after enumerating major risks,agencies should provide model

84、 processes for risk mitigation.Parallel to this,agencies such as the National Science Foundation(NSF)and the National Institute of Standards and Technology(NIST)should continue supporting research in the scientific foundations of responsible and trustworthy AI.This research should include the develo

85、pment of standard benchmarks to measure AI model properties such as accuracy,reproducibility,fairness,resilience,and explainable AI,as well as AI algorithms that monitor themselves for these properties and adjust when the benchmarks are not within defined norms.Another goal of such research should b

86、e the development of tools to evaluate biases in data sets and to distinguish synthetic data from real world data.Recommendation 5:Encourage innovative approaches to integrating AI assistance into scientific workflows.The scientific enterprise is an excellent sandbox in which to practice,study,and a

87、ssess new paradigms of collaboration between humans and AI assistants.The objective should not be to maximize the amount of automation,but to allow human researchers to achieve high quality science that utilizes AI assistance responsibly.Funding agencies should recognize the emergence of these new w

88、orkflows and design flexible procedures,metrics,funding models,and challenge problems that encourage strategic 10 National Science Foundation.(Accessed 23 April 2024).The State of U.S.Science and Engineering 2022(Figure 16).11 A foundation model is an ML model trained(often at great computational ex

89、pense)on a broad range of data,which can then be fine-tuned relatively cheaply for more specialized applications.For more discussion,see Bommasani,R.et al.(2022 July).On the Opportunities and Risks of Foundation Models.arXiv.12 National Institute of Standards and Technology.(2023 November).Artificia

90、l Intelligence Safety Institute Consortium.Federal Register.14 experimentation with new AI-assisted ways to organize and execute a scientific project.Implementation of these workflows also present opportunities for researchers from a variety of disciplines,such as human factors and industrial and or

91、ganizational psychology,to advance our knowledge in the area of human-machine teaming.More broadly,incentive structures across funding agencies,academia,and the academic publishing industry may need to be updated to support a broader range of scientific contributions,such as curating a high quality

92、and broadly usable data set,that might not be given sufficient recognition by traditional metrics of research productivity.15 1.Introduction We stand on the brink of an extraordinary era where innovation driven by artificial intelligence(AI)across the sciences promises to greatly accelerate Americas

93、 long-term leadership in scientific knowledge and solutions.13,14,15,16 This AI-powered paradigm shift in scientific tools and methods is timely and critical as humanity confronts daunting challenges in important areas such as energy,climate,healthcare.Beyond solving known challenges,by harnessing A

94、I responsibly,equitably,and effectively for research,scientists can deliver greater resilience to our society,improving our ability to provide a vast array of benefits such as clean water,abundant electricity,and health and wellness to Americans.Yet the most important payoff of the AI transformation

95、 of the sciences will be the ability to realize previously unattainable or even unimagined possibilities with scientific advancements and understandings developments that have the potential to improve the lives of all Americans.We are already seeing glimmers of possibility for leveraging AI to addre

96、ss difficult health challenges,including cancer,17 autoimmune diseases,18 neurodegenerative disorders,19 and drug-resistant infection.20 At the same time,researchers are harnessing AI to generate surprising advances in materials science21 with promise for creating next-generation batteries,22 superc

97、onductors,23,24 and computer chips.25 Beyond advances in core science and engineering disciplines,AI methods promise to provide high fidelity models“digital twins”26of the world that can help us to cut through uncertainty and complexity to predict,to plan,and to guide policymaking,where scarce data

98、and models currently make it difficult to assess potential pathways forward.13 Hope,T.et al.(2023 August).A Computational Inflection for Scientific Discovery.Communications of the ACM.14 Wang,H.et al.(2023 August).Scientific discovery in the age of artificial intelligence,Nature.15 Mock,M.et al.(202

99、3 September).AI can help to speed up drug discovery but only if we give it the right data.Nature.16 National Academies of Sciences,Engineering,and Medicine.(2022 May).Automated Research Workflows for Accelerated Discovery:Closing the Knowledge Discovery Loop.17 Thierry,A.(2023 January).Circulating D

100、NA fragmentomics and cancer screening.Cell Genomics.18 Danieli,M.et al.(2024 February).Machine learning application in autoimmune disease:State of art and future prospects.Autoimmunity Reviews.19 Cumplido-Mayoral,I.et al.(2023 April).Biological brain age prediction using machine learning on structur

101、al neuroimaging data:Multi-cohort validation against biomarkers of Alzheimers disease and neurodegeneration stratified by sex.eLife.20 Wong,F.et al.(2023 December).Discovery of a structural class of antibiotics with explainable deep learning.Nature.21 See section 3.1 of this report.22 Chen,C.et al.(

102、2024 January).Accelerating computational materials discovery with artificial intelligence and cloud high-performance computing:from large-scale screening to experimental validation.arXiv.23 Pogue,E.et al.(2023 October).Closed-loop superconducting materials discovery.npj Computational Materials 24 Li

103、u,Y.et al.(2023 December)Materials Expert-Artificial Intelligence for Materials Discovery.arXiv.25 E.g.,Liu,M.et al.(2024 April).ChipNeMo:Domain-Adapted LLMs for Chip Design.arXiv.26 A digital twin is a high-resolution model of a physical system that is continually updated with real-time data from t

104、hat system.Such twins usually rely on traditional simulation to model the fundamental processes of the system but can additionally use AI models to refine,accelerate,or analyze such simulations.For the current state of such models,see National Academies of Science,Engineering,and Medicine.(2024).Fou

105、ndational Research Gaps and Future Directions for Digital Twins.National Academies Press.16 In the October 2023 Executive Order on the Safe,Secure,Trustworthy Development and Use of Artificial Intelligence,27 PCAST was charged to report on the potential impact of AI on scientific research aimed at t

106、ackling major societal and global challenges,and on practices needed to ensure effective and responsible use of these technologies.The effects of AI on such important topics as national security and critical infrastructure,labor markets,authenticity of content such as images or video,intellectual pr

107、operty rights,education,and the criminal justice system will be addressed in other reports and actions mandated by the Executive Order.As such,this report focuses specifically on the role of AI in the sciences,rather than these broader impacts.An acceleration of science Broadly speaking,scientific a

108、dvances have historically proceeded via a combination of three paradigms:empirical studies and experimentation;scientific theory and mathematical analyses;and numerical experiments and modeling.In recent years a fourth paradigm,28 data-driven discovery,has emerged.These four paradigms complement and

109、 support each other.However,all four scientific modalities experience impediments to progress.Verification of a scientific hypothesis through experimentation,careful observation,or via clinical trial can be slow and expensive.The range of candidate theories to consider can be too vast and complex fo

110、r human scientists to analyze.Truly innovative new hypotheses might only be discovered by fortuitous chance,or by exceptionally insightful researchers.Numerical models can be inaccurate or require enormous amounts of computational resources.Data sets can be incomplete,biased,heterogeneous,or noisy t

111、o analyze using traditional data science methods.AI tools have obvious applications in data-driven science,but it has also been a long-standing aspiration29 to use these technologies to remove,or at least reduce,many of the obstacles encountered in the other three paradigms.With the current advances

112、 in AI,this dream is on the cusp of becoming a reality:candidate solutions to scientific problems are being rapidly identified,complex simulations are being enriched,and robust new ways of analyzing data are being developed.By combining AI with the other three research modes,the rate of scientific p

113、rogress will be greatly accelerated,and researchers will be positioned to meet urgent global challenges in a timely manner.30 Like most technologies,AI is dual use:AI technology can facilitate both beneficial and harmful applications and can cause unintended negative consequences if deployed irrespo

114、nsibly or without expert and ethical human supervision.Nevertheless,PCAST sees great potential for advances in AI to accelerate science and technology for the benefit of society and the planet.In this report,we provide a high-level vision for how AI,if used responsibly,can transform the way that sci

115、ence is done,expand the boundaries of human knowledge,and enable researchers to find solutions to some of societys most pressing problems.We will illustrate this potential for seven different areas of science by 27 Executive Order 14110,88 FR 75191.(2023 October).Safe,Secure,and Trustworthy Developm

116、ent and Use of Artificial Intelligence.28 Data has of course played an important role in the sciences for centuries;but the capability to collect and process such data in the modern era is unprecedented,see Hey,T.et al.(2009 October).The Fourth Paradigm Data-Intensive Scientific Discovery.Microsoft

117、Research.29 Smalheiser,N.and Swanson,D.(1998 November),Using ARROWSMITH:a computer-assisted approach to formulating and assessing scientific hypotheses.Computer Methods and Programs in Biomedicine.30 Wang,H.et al.(2023 August).Scientific discovery in the age of artificial intelligence,Nature.17 desc

118、ribing the current state-of-the-art in the field,the ways that AI is beginning to be deployed to overcome barriers to progress,the benefits we think AI can help to achieve in the mid-to-long-term,and examples of challenges that must be overcome.Finally,we provide cross-cutting findings and recommend

119、ations about what will be needed to realize our vision for the future of AI-enabled science while mitigating potential risks.Artificial intelligence and machine learning AI refers to a wide spectrum of technologies intended to perform or assist with cognitive tasks that were previously only achievab

120、le via human intelligence.AI tools do not actually duplicate the mechanisms of human thought,which are themselves still incompletely understood by modern science.Rather,most AI systems currently operate through machine learning(ML),which is an array of techniques that leverages statistical inference

121、“learned”through training of an AI model on large data sets.31 Such models may then be applied to infer answers to related questions.AI systems do not have a deep conceptual comprehension of the task that they are attempting to solve,32,33 and in the absence of external verification,the answers prov

122、ided by an AI model are not guaranteed to be correct.Nevertheless,these models can perform remarkably well at many complex and imprecisely specified tasks,using their training process to identify patterns and relationships that were previously hidden in the data sets.Over the last twenty years,numer

123、ous ML technologies have matured to the point where they are routinely deployed in consumer and business products.As classic examples,ML powers filters for junk email and is also used to prioritize email.In more recent applications,a commercially available smartphone is capable of recognizing speech

124、 commands,identifying its owner through facial recognition,34 or extracting text from an image and then translating it to a different language.35 ML technology is also found in cars:sensors,self-piloting features,and other driver assistance systems are being deployed in newer vehicles to significant

125、ly improve road safety.36 Today,many companies routinely rely on ML algorithms to detect fraud,improve their logistics and marketing,deliver targeted advertisements to consumers,or predict customer creditworthiness.Despite these proven applications,ML and broader AI technologies still suffer from se

126、veral weaknesses.Their outputs are often arrived at by an opaque process that carries no guarantee of correctness and which may involve usage of data that is either protected by intellectual property rights or contains sensitive private information about individuals.Biases in the training data set,a

127、s well as systematic biases in the training process,can lead to problematic algorithmic bias in the 31 Several AI systems also have an additional fine-tuning component to their training,for instance by adjusting the model parameters in response to human responses to its outputs by a process known as

128、 reinforcement learning with human feedback.32 Indeed,there is not even a consensus on what it means for an AI to understand a concept.33 Joksimovic,S.et al.(2023).Opportunities of artificial intelligence for supporting complex problem-solving:Findings from a scoping review.Computers and Education:A

129、rtificial Intelligence.34 E.g.,Idemia Facial Recognition Access Control is now deployed at many airports worldwide to streamline security procedures.35 Popular software tools in this category include the Voice Dream Scanner for low sighted persons,or Google Translate.36 According to a recent study b

130、y the AAA,such systems could prevent as many as 37 million automotive crashes over the next 30 years.18 model outputs.To address these issues,the Biden Administrations Blueprint for an AI Bill of Rights37 set out guiding principles on how to mitigate algorithmic bias and other weaknesses prior to in

131、corporation into real world systems.The NIST AI Risk Management Framework38 added further specific measurement and managerial approaches to mitigate and reduce adverse outcomes from adoption of AI,and the Office of Management and Budget(OMB)Memo on Advancing Governance,Innovation,and Risk Management

132、 for Agency Use of Artificial Intelligence39 further clarified how federal agencies should utilize AI.Generative AI Some of the most prominent and striking recent advances in the field have occurred in the class of machine learning tools known as generative AI,which include popular large language mo

133、dels such as ChatGPT,Gemini,Claude,and LLaMA,image generation models such as Midjourney and DALL-E,and scientific generative models such as AlphaFold,RoseTTAfold,and ChemBERTa.By using large and complex deep neural network40 models which have been trained on many terabytes41 of data,and which have o

134、ften been refined by thousands of hours of human input,these generative AI tools analyze user prompts and produce outputs using models that have billions or even trillions of learned relationships or parameters.Generative algorithms extend inputs into likely sequences or structures,enabling these mo

135、dels to create text,images,and other media in response to prompts that many liken to the responses of a human expert.Reinforcement learning from human feedback can further enable improvement of these generative models.Generative AI is still a very new and rapidly developing technology,and as such ha

136、s issues and problems that warrant further research and improvement.For instance,in addition to the general weaknesses of machine learning tools previously mentioned,some generative AI outputs are also prone to“hallucination.”Since generative AI models do not exclusively encode factual knowledge,the

137、y can confidently make assertions that have no factual basis but are deceptively convincing,or may produce images with artifacts that are inconsistent with physical reality.The challenge with the veracity of AI generations extends beyond the fidelity of training data:the probabilistic nature of gene

138、rations may lead AI models to create plausible,yet inaccurate generations.In some applications,e.g.,working with AI systems to generate out-of-the-box possibilities,such“imaginative”processes can be useful or desired.However,in most scientific applications,the goal is truthful inferences.Strategies

139、and mitigations continue to be developed to reduce instances of erroneous generations and other observed problems with the output of generative AI.Methods are being investigated to determine when scientific AI models generate errors,42 to provide a well-calibrated confidence level in the output of s

140、uch models,43 and to ensure that such outputs are consistent with physical,37 The White House.(2022 October).Blueprint for an AI Bill of Rights:Making Automated Systems Work for The American People.38 National Institute of Standards and Technology.(2023 January).Artificial Intelligence Risk Manageme

141、nt Framework(AI RMF 1.0).39 The White House Office of Management and Budget.(2024 March).Advancing Governance,Innovation,and Risk Management for Agency Use of Artificial Intelligence.Memo M-24-10.40 Deep neural networks are a particularly successful class of machine learning algorithms that were ins

142、pired by the structure of neurons inside a human brain.Neural networks are composed of interconnected layers each containing a simpler unit or“node”that can conducts part of a computation.41 E.g.,of data volumes:University of Delaware.Examples of Data Volumes.(Accessed 2024 April 16).42 Azaria,A.and

143、 Mitchell,T.(2023 October).The Internal State of an LLM Knows When Its Lying.arXiv.43 Nori,H.et al.(2023 April).Capabilities of GPT-4 on Medical Challenge Problems.arXiv.19 biological,or chemical constraints.44 There are further concerns with generative AI technology.The most powerful models current

144、ly require large amounts of computational infrastructure and energy45 to train and fine-tune.Furthermore,most of these models are constructed and controlled by a small number of large technology companies,with the current leaders in the field mostly based in the United States.This situation continue

145、s to evolve.For instance,reasonably capable open source and/or open weight46 generative models have been released,and recent research has demonstrated that small language models can provide high levels of performanceat times comparable to the output of the largest modelswith significantly lower reso

146、urce costs.47 While AI tools have advanced considerably,the prospect of employing generative AI to fully replace humans in real world processes and workflows is still largely in the realm of the far future,and the future this report envisions is one in which AI tools assist rather than supplant huma

147、ns.Despite their imperfect nature,generative AI technologies hold tremendous potential and promise.According to one recent estimate,applications of generative AI could add between$2.6 trillion to$4.4 trillion annually to the global economy.48 AI technologies present especially transformative opportu

148、nities in the realm of scientific research,in which many of the aforementioned weaknesses can be addressed by combining AI models with other scientific tools and methods.There is a growing body of research demonstrating that external scientific validation methods,such as laboratory experiments,clini

149、cal trials,numerical simulations,formal verification software,and limiting the AI input to draw from curated scientific literature to draw from curated scientific literature can be used to mitigate the hallucinations of AI models.For example,rather than being trained on large“internet scale”text res

150、ources,scientific models can be trained on extensive scientific data sets that are publicly available and better curated,and which come with relatively few intellectual property restrictions.The scientific community places great value on transparency and reproducibility,and many AI models and method

151、ologies used in research are openly available.The sharing of data and training resources can greatly reduce the considerable total cost in time,computational resources,and energy consumption needed for scientists to build and train these models.Sharing resources also facilitates validation of output

152、s,verification of results,and building directly on prior work.Basic research can also be a platform for experimental trials of new AI technologies,which could yield valuable insights for how(or whether)to responsibly deploy AI in other contexts.The advancing capabilities of AI technologies in the sc

153、iences does introduce new potential risks,particularly concerning malicious applications.The enhanced capabilities also present novel challenges to and pressures on the scientific communitysuch as the risk of placing too much trust 44 For a discussion of physics-informed machine learning methods in

154、general,see Karniadakis,G.E.et al.(2021 May).Physics-informed machine learning.Nature Reviews Physics.45 Additionally,the cooling requirements of the data centers used by AI can lead to significant water usage.46 Open weight models disclose the weights obtained at the end of the training process,whe

155、reas open source models disclose the structure of the model and training process,but not necessarily the final weights.This distinction is particularly significant for generative pretrained transformer(GPT)models,for which the training process is extremely computationally intensive.47 Li,Y.et al.(20

156、23 September).Textbooks Are All You Need II:Phi-1.5 Technical Report.arXiv.48 Chui,M.et al.(2023 June).The economic potential of generative AI:The next productivity frontier.McKinsey Digital.20 on a convenient AI tool and not checking the resultswhich could impact established norms and principles of

157、 scientific integrity.Nonetheless,if AI tools are deployed in a responsible,human-supervised,and validated manner,the scientific benefits of these technologies can outweigh the risks.Indeed,since our global competitors and collaborators will certainly be developing applications of AI,the best way to

158、 mitigate its risks is to lead in the development of norms and best practices.The greater risk would be not seizing the opportunity to lead the world on developing and understanding these powerful tools,and applying them to our most pressing global problems.2.A Vision for the Future of AI-Enabled R&

159、D Given the experimental and rapidly developing nature of scientific applications of AI technologies(particularly with regards to generative AI),it is difficult to make predictions about how and when the scientific workflow will evolve as AI techniques are incorporated.Nevertheless,we envision that

160、the ideal future of AI-enabled science will require continued attention in three areas:Empowerment of human scientists;Responsible use of AI tools;and Sharing of basic AI resources.As depicted in the accompanying figure,these three themes mutually reinforce each other.Sharing basic AI infrastructure

161、,like access to time on high performance computing clusters,will enable scientists to work with and understand advanced AI tools,help set common standards for responsible use,and improve equity by providing access to researchers from all institutions rather than just those with the most resources.A

162、culture of responsible use will encourage secure ways to share models and data,as well as promote thoughtful designs of AI-assisted research projects that enhance,rather than degrade,the quality of their scientific output.Finally,a diverse scientific Figure 1.Virtuous cycle of responsible use,shared

163、 resources,and empowerment to leverage AI to accelerate scientific research.Multiple dimensions of this interplay are explored in this report.21 community that is broadly empowered by AI tools will generate innovative new solutions to pressing challenges,justifying the initial investments in shared

164、resources and in responsible use policies,as well as creating a community of stakeholders to continue to develop and build upon these investments.Our recommendations at the end of this report are designed to support and encourage all three of these themes.AI methods will help researchers prioritize

165、the most likely solutions Many of the problems at the frontier of modern science are complex and interdisciplinary.As such it is becoming increasingly difficult for unaided human experts to sift through the large amounts of available data and analysis from all relevant scientific domains or to evalu

166、ate hundreds of thousands of candidates(e.g.,compounds for medicine or materials for engineering applications)to identify the most promising solutions to a given problem.49 Scientists are beginning to use machine learning tools,including generative AI,to address this gap.Researchers are now using da

167、ta-driven models to isolate likely candidates for materials,50 drugs,51 and chip designs52 to test in a laboratory or clinical trialpotentially saving enormous amounts of time and expense by reducing the number of tests with negative outcomes and leveraging the most value out of limited experimental

168、 resources.In the future,these tools could also suggest possible explanations for empirically observed phenomena,or uncover connections or analogies between two areas of science that would otherwise have gone unnoticed.Hypothetically,AI could even help scientists to discover new laws of nature,which

169、 could then be validated by more traditional theoretical calculations and laboratory experiments.It is important to be clear that we do not envision that AI-driven or AI-assisted reasoning will replace the uniquely human capabilities and genius to make inspired connections and conclusions.Rather,we

170、expect that traditional forms of research will continue to play an essential role in the broader scientific enterprise for the foreseeable future.By handling routine tasks,AI will allow scientists to focus on core research Some of the most immediate productivity gains offered by AI will come not fro

171、m directly assisting with the most difficult scientific research challenges,but rather from the more mundane support AI can offer with secondary tasks that can take up a large portion of the working time of a scientist,such as developing computer code,assisting with papers and reports,53 performing

172、literature reviews,and acquiring expertise in adjacent scientific fields.Already,general purpose large language models are 49 Bloom,N.et al.(2020 April).Are Ideas Getting Harder to Find?.American Economic Review.50 Zeni,C.et al.(2024 January).MatterGen:a generative model for inorganic materials desi

173、gn.arXiv;See also the A-Lab and the Materials Project,both operated by Berkeley Lab.51 Mock,M.et al.(2023 September).AI can help to speed up drug discoverybut only if we give it the right data.Nature.52 E.g.,Liu,M.et al.(2024 April).ChipNeMo:Domain-Adapted LLMs for Chip Design.arXiv.53 Currently,the

174、 output from existing large language models is not of sufficiently reliable quality to be acceptable for direct use in writing scientific documents,though it can already serve to create useful first drafts or experimental variants of such texts.Nevertheless,we expect more professional quality AI wri

175、ting assistants to be integrated into many text editing platforms in the near future,and to eventually gain substantial adoption within the scientific community,in parallel with an updating of professional writing standards and guidelines that takes into account the capabilities and limitations of s

176、uch AI assistants,for instance through the development of benchmark datasets for scientific writing,similar to recent benchmarks such as LegalBench in the practice of law.Furthermore,human authors should be held accountable for any errors in AI-generated writing output,and be expected to maintain fu

177、ll editorial control of content.22 used routinely.54,55,56,57,58 Nevertheless,we expect hands-on,non-AI-assisted research experiences to remain an essential component of the training of junior scientists for the foreseeable future.Rote laboratory processes will be automated and improved Many routine

178、 laboratory processes are ideal for labor saving with AI,allowing humans to spend their time and energy on the things we do best and find most interesting,such as design,analysis,and collaboration.A number of AI technologies,including object recognition,reinforcement learning,59 and generative AI,ar

179、e beginning to be used to control robotic systems,allowing them to process complex instructions in unpredictable environments and adapt to sensory feedback.60 This flexible problem-solving capability will soon allow AI-powered robots to be used in the laboratory to perform a wide range of experiment

180、s,or synthesize a large number of materials,without the need to carefully reprogram the robot with each new task.Virtually every aspect of the laboratory workflow,from experimental design to data collection to data interpretation,could be partially or fully automated through AI assistance,61 althoug

181、h we view expert human supervision of such automated laboratories to be essential and highly desirable for decades to come.Previously intractable simulations will become possible AI algorithms are successfully being used to greatly accelerate and enhance the computationally expensive computer models

182、 used to simulate complex systems,such as Earths climate,the quantum chemistry of materials,and the intricate dynamics of proteins and cellular structures,reducing the need to always return to time-consuming modeling of these systems from first principles.For example,AI models are being used to prov

183、ide more efficient approximations for solving the Schrdinger equation for chemical compounds,the solutions of which help to define compound 54 Ray,P.(2023 April).ChatGPT:A comprehensive review on background,applications,key challenges,bias,ethics,limitations and future scope.Internet of Things and C

184、yber-Physical Systems.55 Huang,J.and Tan,M.(2023 April).The role of ChatGPT in scientific communication:writing better scientific review articles.American Journal of Cancer Research.56 Boiko,D.et al.(2023 December).Autonomous chemical research with large language models.Nature.57 It is however possi

185、ble that a version of the Jevons paradox(or the myth of the paperless office)will take hold,namely that the increased efficiency of writing afforded by AI tools is converted(in the near term,at least)into a greater amount of scientific writing being produced,rather than a reduction in time spent on

186、writing tasks.58 Sellen,A.and Horvitz,E.(2023 November).The Rise of the AI Co-Pilot:Lessons for Design from Aviation and Beyond.arXiv.59 Reinforcement learning is a method of training algorithms to make desired actions by requiring the process to maximize a given function or result.60 Zhou,C.et al.(

187、2021 November).A review of motion planning algorithms for intelligent robots.Journal of Intelligent Manufacturing.61 National Academies of Sciences,Engineering,and Medicine.(2022 May).Automated Research Workflows for Accelerated Discovery:Closing the Knowledge Discovery Loop.23 stability and other p

188、roperties.62 Multiple efforts are now underway63,64,65,66,67 to develop AI-powered“foundation models”68 and“digital twins”for many applications.When complete,entire communities of researchers will be able to build on these powerful and broad platforms to rapidly create more customized models for a w

189、ide variety of scientific and engineering purposes.“Lightweight”versions of large models will also be developedthese will be small enough to run and fine-tune on individual computers,while still retaining much of the capability of the original model.69 This will reduce the cost and environmental imp

190、act of deploying AI in scientific applications,and should enable wider,more equitable access to the models and resulting innovations.As previously noted,care should be taken to avoid reliance purely on AI models.Whenever possible,AI simulations should be validated and benchmarked against traditional

191、 methods of verification such as numerical simulations,theoretical calculation,and agreement with experimental data.Shared models and data will reduce duplication of effort,democratize research,and reduce the total cost of using AI For AI tools to be used effectively,high quality data sets must firs

192、t be collected,made accessible,and put into a usable format.Next,models need to be trained on that curated data,taking care to limit algorithmic bias while also protecting the privacy of any humans whose personal information might be part of the data set.70,71 Repeating these steps for each individu

193、al research project is inefficient,wasting considerable time and resources.For example,instead of the multiple efforts currently underway to build large-scale foundation models for biology,it is feasible that these efforts could be combined or connected to focus resources and talent on model scale a

194、nd quality.Through central resources such as the envisioned National Artificial Intelligence Research Resource(NAIRR),72 researchers will gain access to standardized models and curated data sets,and share best practices 62 Radu,A.and Duque,C.(2022 February).Neural network approaches for solving Schr

195、dinger equation in arbitrary quantum wells.Scientific Reports.63 E.g.,Mukkavilli,S.et al.(2023 September).AI Foundation Models for Weather and Climate:Applications,Design,and Implementation.arXiv.64 E.g.,Houben,M.(2020 November).Digital Twins,the future in plant phenotyping TechnoHouse by Rijk Zwaan

196、.Phenospex.65 E.g.,Costin,A.et al.(2023 November).Digital Twin Framework for Bridge Structural Health Monitoring Utilizing Existing Technologies:New Paradigm for Enhanced Management,Operation,and Maintenance.Transportation Research Record:Journal of the Transportation Research Board.66 E.g.,van Will

197、egen,B.et al.(2022 September).A review study of fetal circulatory models to develop a digital twin of a fetus in a perinatal life support system.Frontiers in Pediatrics.67 E.g.,Geddes,L.(2023 November).How digital twins may enable personalised health treatment.The Guardian.68 A foundation model is a

198、n ML model trained(often at great computational expense)on a broad range of data,which can then be fine-tuned relatively cheaply for more specialized applications.For more discussion,see Bommasani,R.et al.(2022 July).On the Opportunities and Risks of Foundation Models.arXiv.69 Javaheripi,M.and Bubec

199、k,S.(2023 December).Phi-2:The surprising power of small language models.Microsoft Research Blog.70 Dilmaghani,S.et al.(2019 December).Privacy and Security of Big Data in AI Systems:A Research and Standards Perspective.2019 IEEE International Conference on Big Data,IEEE Xplore.71 Metcalf,J.and Crawfo

200、rd,K.(2016 June).Where are human subjects in Big Data research?The emerging ethics divide.Big Data&Society.72 The National Artificial Intelligence Research Resource(NAIRR)Pilot.(Accessed 2024 April 9).24 that will advance the fundamental science of AI.73 A central resource would also facilitate the

201、installation,fine-tuning,and operation of the aforementioned foundation models by researchers in a broad variety of domains without requiring significant levels of specialized AI expertise.Once the right foundational infrastructure74 is in place,AI-assisted research will become possible not only for

202、 highly resourced companies and research groups,but also smaller institutions and private sector organizations,or even members of the general public,75 creating more equitable opportunities for discovering and developing innovative ways to utilize AI tools.However,for models and data that could have

203、 potentially harmful applications or which require privacy protections,some restrictions on access will be needed.The U.S.government is taking steps to facilitate safe and trustworthy AI.For instance,the recently released OMB memo outlines a set of minimum risk practices that federal agencies will b

204、e mandated to follow to mitigate risks of AI in rights-and safety-impacting contexts.76 The NAIRR pilot is providing access to data,training,and compute resources to researchers who are conducting research on trustworthy machine learning.In addition,the recently formed U.S.Artificial Intelligence Sa

205、fety Institute and Artificial Intelligence Safety Institute Consortium(AISIC)will bring together experts from across government,academia,and civil science to collaborate on AI safety research and develop resources.Multimodal foundation models will bring together multiple forms of data and create new

206、 synergies among branches of science Large-scale language models have been referred to as“foundation models”because they provide a basis or“foundation”for efficient refinement that transforms general AI models into AI systems for specialized tasks and domain-specific applications.Such refinement is

207、usually done via learning mechanisms that use specialized,domain-specific data.The process of refining foundation models into high-performance models in specific domains is a type of transfer learning.77 In machine learning research,this term has been used for decades to refer to methods involving t

208、he adaptation of an ML model trained on one domain,e.g.,inorganic chemistry,to harness its“pretraining”to improve the models ability to learn in other domains,such as predicting the function of proteins and cells.73 National Artificial Intelligence Research Resource Task Force.(2023 January).Strengt

209、hening and Democratizing the U.S.Artificial Intelligence Innovation Ecosystem:An Implementation Plan for a National Artificial Intelligence Research Resource.74 Such infrastructure includes not only hardware and software resources,but also benchmarks,regulatory principles,and guidelines on the respo

210、nsible usage of AI(such as the IRB Considerations on the Use of Artificial Intelligence in Human Subjects Research);and,most importantly,the human capital of experts in the use and deployment of AI in both public and private sectors.75 For certain dual use scientific applications,such as gain of fun

211、ction research for viruses,broadening access of AI-assisted scientific advances to the general public is not necessarily desirable,and will require some additional regulation and oversight.However,we see many areas of science where it would be beneficial to have increased participation and engagemen

212、t with the public,and to not have the most powerful AI tools limited to only a small number of well-resourced groups,or to researchers based in other countries.76 The White House Office of Management and Budget.(2024 March).Advancing Governance,Innovation,and Risk Management for Agency Use of Artifi

213、cial Intelligence.Memo M-24-10.77 E.g.,Zhuang,F.et al.(2021 January).A Comprehensive Survey on Transfer Learning.Proceedings of the IEEE,IEEE Xplore.25 In recent work,researchers have been exploring the construction of multiscale78 and multimodal foundation models that can take advantage of joint re

214、presentations79 learning as well as harnessing transfer learning.The aim of multimodal learning is to combine a diverse array of data sets,including data of different types,of different scales,and even from different fields of science.80,81 Beyond learning jointly from datasets containing multiple t

215、ypes of data,efficient methods are being developed for joining models that have been trained independently with data drawn from different scientific areas or foci.Work in this realm includes the use of adaptors:lightweight connector models that are trained to link two or more pre-existing models tog

216、ether.While multimodal models and associated capabilities can be constructed in different ways,we anticipate that developing AI tools that span multiple disciplinary realms and multiple spatial and temporal resolutions82 will provide scientists with powerful emergent capabilities to describe or simu

217、late complex systems.These capabilities may greatly exceed what can be accomplished by domain-specific models alone,and will open up rich new opportunities for interdisciplinary thinking and collaboration.AI will help researchers do more with data Through the demonstrated ability of AI to infer cont

218、ext and make use of complex,noisy,non-quantitative real-world data,such as natural language text,AI algorithms show great potential for automatically organizing,combining,and“cleaning”83 the extremely large and heterogeneous data sets that underpin modern data-driven science,as well as identifying a

219、nomalies and uncovering important correlations and patterns within that data.AI tools are also providing new ways to achieve superresolution to enhance the quality of individual images.84 In addition,in many fields AI tools are already being used routinely to help generate“synthetic”data sets thatwh

220、en used responsiblycan greatly enhance the quality and predictive power of empirically generated data,protect the privacy of sensitive information in such data sets,reduce the risk of algorithmic bias,and help extrapolate from the underlying data to draw conclusions in broader domains.85,86,87,88 Ho

221、wever,such synthetic data will need to be carefully and permanently labeled as distinct from data 78 Multiscale modeling is a modeling strategy that uses multiple models at different scales simultaneously to describe a system.79 Joint representations are a means for machine learning models to learn

222、from multiple types of data or features,such as images combined with text,in a unified manner.80 Fei,N.et al.(2022 June).Towards artificial general intelligence via a multimodal foundation model.Nature Communications.81 Li,Z.et al.(2024 February).MLIP:Enhancing Medical Visual Representation with Div

223、ergence Encoder and Knowledge-guided Contrastive Learning is a recent example of a multimodal model that learns from both the image and text of annotated medical images.82 Poli,M.et al.(2023 March).Ideal Abstractions for Decision-Focused Learning.arXiv.83 Data cleaning refers to the process of remov

224、ing or repairing portions of a data set that are duplicated,incomplete,inaccurate,or unrelated,in order to improve the quality of that data set for analysis or training.84 E.g.,Chen,H.et al.(2022 March).Real-world single image super-resolution:A brief review.Information Fusion.85 Laboratory for Info

225、rmation and Decision Systems.(2020 October).The real promise of synthetic data.MIT News.86 Savage,N.(2023 April).Synthetic data could be better than real data.Nature Outlook:Robotics and Artificial Intelligence.87 Jordon,J.et al.(2022 May).Synthetic Datawhat,why and how?arXiv.88 Listgarten J.(2024 J

226、anuary).The perpetual motion machine of AI-generated data and the distraction of ChatGPT as a scientist.Nature Biotechnology.26 collected from real world observations,sensing,surveys,and experiments,to avoid contributing to the longer term problem of data pollution.89 New forms of collaboration will

227、 emerge A modern scientific research project is typically led by a small number of senior scientists directing a larger group of postgraduate researchers and students to perform more specialized subtasks.AI tools will automate,or at least assist with,many of these subtasks,and through their facility

228、 with natural language,greatly facilitate the way that researchers from different scientific backgrounds,levels of expertise,or primary languages can communicate with each other.Because of this,new paradigms for collaboration will emerge,such as AI-augmented experts,human-advised AI systems,hybrid A

229、I techniques combining complementary AI technologies,90 or large crowdsourced,decentralized,and/or highly interdisciplinary projects in which individual contributions are validated and collated through a combination of AI tools and more rigorous assessment methods.91 Beyond facilitating communicatio

230、ns between scientists,AI assistance could allow lay individuals to provide input to cutting-edge research projects.For instance,the public could have opportunities to contribute directly and meaningfully to research in novel ways via specialized chatbots that could engage on demand in a scientifical

231、ly accurate,accessible,and genuinely two-way fashion in which public comments fold back into hypotheses or formulation of outputs.92 Even in a future where AI assistance becomes commonplace,we envision a continuing role for traditional forms of scientific research and engagement.Traditional research

232、 methods and approaches are an essential and complementary approach to AI-assisted science,providing qualitative context,vital proof through experimentation,crucial methods for training scientists so they understand the mistakes that can come from tools like AI,and,perhaps most importantly,supplying

233、 the intellectual diversity of the scientific enterprise that refines scientific thought.Responsible AI practices will be integrated into research workflows The scientific method inherently incorporates self-correction mechanisms through independent replication and review.In particular,the design of

234、 experiments has continuously evolved as science has matured,with standard procedures introduced over time to reduce bias,minimize harm to human or animal subjects,improve replicability,and avoid waste.A similar evolution will take place with the use of AI tools in researchbut it will require that a

235、cademic institutions,scientific professional organizations,and funding agencies commit to building a culture of responsible and ethical AI practice,93 and help to make powerful models available to researchers to develop,test,and evaluate these practices.Scientific,academic,and government leaders mus

236、t encourage best practices,such as scientists ensuring appropriate citations of AI model usage in their work and results,and sharing with their communities the details about the specific version of model used;these citations will be critical for facilitating replication and verification of results.P

237、CAST also 89 Ben-Shahar,O.(2019 September).Data Pollution.Journal of Legal Analysis.90 Marcus,G.(2020 February).The Next Decade in AI:Four Steps Towards Robust Artificial Intelligence.arXiv.91 For some emerging examples of such large-scale collaborations in the mathematical sciences,see Artificial I

238、ntelligence to Assist Mathematical Reasoning:Proceedings of a Workshop,National Academies of Sciences,Engineering,and Medicine.(2023 June).92 E.g.,PCAST.(2023 August).Advancing Public Engagement with the Sciences.93 E.g.,The Hastings Center.(2024 April).AI Code of Conduct Draft is Released:Submit Yo

239、ur Comments.27 anticipates advances in the development of attribution tools and affordances94 that can provide scientists with an understanding of the sources in the training data,which is fundamental to the generation of output,so that prior research and results can be cited appropriately.Through a

240、dvances in the foundational computer science of AI,metrics are being developed to determine the quality and optimize the choice of AI training data sets,along with standardized methods to compensate for biases and omissions in such sets.PCAST expects that new AI architectures will be developed95 tha

241、t can offer comparable or superior performance to the largest existing models but at a fraction of their energy consumption and environmental impact.Sophisticated privacy protection methods96,97,98 are beginning to be deployed to protect the privacy of sensitive personal data,such as medical informa

242、tion,that could otherwise leak into and out of an AI model.Hallucinations and other inaccuracies in AI models can be compensated for by refining methods that can ground AI outputs with real world sources,and to attach well-calibrated confidence levels to those outputs;we can also harness more rigoro

243、us validation methods,such as formal verification.99 Models,data sets,and weights should be responsibly open sourced when feasible,to assist with replication and transparency,100 with a preference for explainable AI models that can draw explicit links between their conclusions and their input data.1

244、01 Best practices for maintaining human supervision of automated laboratory processes should also be established.Through increased dialogue between the physical sciences,social sciences,and the humanities,we expect to achieve a greater understanding of the potential unintended consequences of deploy

245、ing AI tools in research,and to develop standard processes to assess,prevent,minimize,and mitigate such consequences.Conversely,while there are significant concerns that the scientific literature is at risk of experiencing an influx of low-quality articles that are partially or even fully AI-generat

246、ed,we also expect AI tools to be useful in upholding standards of academic conduct,for instance by detecting image manipulation,plagiarism,and missing citations.Once the necessary AI infrastructure is in place,new scientific moonshots will become possible When foundational resources such as the NAIR

247、R are in place to provide access to computational power,secure data sharing services,open source(and open weight)AI models,and other key infrastructure,it will become possible to plan ambitious,complex,and large-scale moonshot 94 The uses or purposes that a thing can have;the qualities or properties

248、 of an object that define its possible uses or make clear how it can or should be used.95 One potential future architecture is that of neuromorphic computing;see,e.g.,Schuman et al.(2022 January).Opportunities for neuromorphic computing algorithms and applications.Nature Computational Science.96 E.g

249、.,Papernot,N.and Thakurta,A.(2021 December).How to deploy machine learning with differential privacy.NIST Cybersecurity Insights.97 Savage,N.(2023 April).Synthetic data could be better than real data.Nature Outlook:Robotics and artificial intelligence.98 Kaissis,G.et al.(2020 June).Secure,privacy-pr

250、eserving and federated machine learning in medical imaging.Nature Machine Intelligence.99 Urban,C.and Min,A.(2021 April).A Review of Formal Methods applied to Machine Learning.arXiv.100 National Academies of Sciences,Engineering,and Medicine.(2018 July).Open Science by Design:Realizing a Vision for

251、21st Century Research.101 E.g.,see Xu,F.et al.(2019 September).Explainable AI:A Brief Survey on History,Research Areas,Approaches and Challenges.Natural Language Processing and Chinese Computing.28 scientific projects involving multiple public and private partners.Examples of such projects could inc

252、lude a foundation model to simulate the complexities of the human cell,allowing for in silico(rather than in vitro or in vivo)study of diseases and experimental treatments;a detailed whole Earth model that uses both conventional and AI models to describe components of the Earth system while also bei

253、ng continually updated with highly multimodal real-time data;or a large collaborative effort to discover practical room-temperature superconductors through systematic collection,processing,and AI-assisted analysis of existing data and literature,together with automated laboratory synthesis and testi

254、ng of viable candidates.While there are existing projects along these lines already being undertaken by individual research groups or organizations,the larger scale collaborations that could become possible with an infrastructure of shared AI resources would benefit substantially from economies of s

255、cale and simultaneously reduce duplication of effort.In the next section,we highlight some of these exciting opportunities for advances in discovery that will be made possible using AI.3.Key Opportunities for AI to Supercharge Discovery and Address Global and Societal Challenges The applications of

256、AI tools in science are extremely diverse.In order to illustrate the potential for AI tools to address global and societal challenges,we describe seven representative examples of areas of scientific inquiry to serve as concrete case studies.Each of these fields has faced substantial barriers to prog

257、ress,which,if overcome,could lead to discoveries that improve peoples lives,mitigate global risks,or inspire us as humans.For each of these areas,we characterize the current state of the field,identify barriers to progress,provide examples of how AI is currently being leveraged,and describe the futu

258、re we envision to be possible,along with potential risks that must be considered and resources that would be necessary for progress to be made.This list of examples is not intended to be comprehensive;there are many other fields of science beyond those mentioned here that are also likely to be trans

259、formed by AI.However,these vignettes illustrate the cross-cutting themes,broad range of possibilities,and critical needs that AI offers across the sciences.3.1.A Phase Change for Materials Discovery The use of generative AI for the discovery of new materials102 is already proving to be crucial for d

260、riving economic expansion,improving health outcomes,and creating new materials critical for national defense.Historically,eras of major improvement in the human condition were powered by advances in materials science:bronze,iron,concrete,steel.Today,we live in an age of silicon,hydrocarbons,and nitr

261、ates.The near future may be an era of nanomaterials,103 biopolymers,104 and quantum materials.105 Novel materials will be the basis for climate-friendly energy technologies,including improved batteries and energy storage,carbon capture,and hydrogen production.New biologics will improve health outcom

262、es and open new pathways for care.Finally,AI-assisted R&D will open possibilities that previously only existed in the realm of imagination,such as room temperature superconductors or large-scale quantum computer architectures.Our future will be built on new materials;they will revolutionize our soci

263、ety.102 Snyder,A.(2023 December).An AI boost for developing new materials.Axios.103 PCAST(2023 August).The Seventh Assessment of the National Nanotechnology Initiative.104 PCAST(2022 December).Biomanufacturing to Advance the Bioeconomy.105 Stanev,V.et al.(2021 October).Artificial intelligence for se

264、arch and discovery of quantum materials.Communications Materials.29 Scientists have already had success in leveraging deep learning models for materials discovery.106,107 For example,interdisciplinary teams of researchers at private companies have used AI to develop designs for millions of novel mat

265、erials,where nearly a half a million of those predicted were candidates likely to be stable enough for possible growth in the lab.108 PCAST expects future advances in AI-assisted materials discovery to be capable of narrowing similarly large feature spaces109 of candidate stable materials to isolate

266、 those that are most likely to achieve specified target properties.110 AI has also been used to improve on existing materials;for example,AI tools can help scientists optimize material composition to reduce or eliminate potentially environmentally hazardous materials while maintaining performance.In

267、 addition,AI can incorporate processing parameters to optimize not only composition but also methods of material manufacturing that can increase efficiency,reduce waste,and lead to new reaction and processing pathways that are more sustainable.NSF has made a$72.5 million investment to drive the desi

268、gn,discovery and development of advanced materials needed to address major societal challenges.111 The Designing Materials to Revolutionize and Engineer our Future(DMREF)program will fund 37 new four-year projects that work across many directorates of the NSF in science,engineering,computation,inclu

269、ding partnerships with the private sector;each employing deep learning and AI.112 The Department of Energy has funded the Quantum Science Center at$115M over the next five years,which among other things supports AI-assisted discovery and design of quantum materials.113 AI methods could be particular

270、ly valuable in the search for new superconductors,materials which can conduct electricity so efficiently that there is little to no loss of energy.Superconductors are essential components for MRI machines,particle accelerators,certain experimental quantum computing technologies,and(in limited places

271、)in our power grid for lossless energy transmission;but they currently have several undesirable features.First,all known practical superconductors must be cooled to cryogenic temperatures(-298 degrees Fahrenheit or less)using liquid helium,an impractical procedure involving an expensive and limited

272、resource.Second,in contrast to conventional conductors such as copper,existing superconductors are not malleable and will lose their superconducting properties with damage.Third,they are very expensive to use,both in terms of the cost of their precursor materials and the effort required to engineer

273、them into wires.Better superconductorsthose that can work at more easily achievable temperatures,are easier to engineer into applications,and are cheaperwould be transformative.They would democratize access to MRI,lower energy costs through reduction of resistive losses in our power grid,and enable

274、further electrification of our economy.These materials would also have applications in our 106 Chen,C.et al.(2024 January).Accelerating computational materials discovery with artificial intelligence and cloud high-performance computing:from large-scale screening to experimental validation.arXiv.107

275、Zeni,C.et al.(2023 December).MatterGen:a generative model for inorganic materials design.Microsoft.108 Merchant,A.et al.(2023 November).Scaling deep learning for materials discovery.Nature.109 Feature spaces are collections of qualities or dimensions that are used to characterize data.110 E.g.,see Z

276、hang,Y.and Kim,E.(2017 May).Quantum Loop Topography for Machine Learning.Physical Review Letters for an example of applying machine learning techniques to identify materials exhibiting a topological quantum phase transition.111 National Science Foundation.(2023 September).NSF invests$72.5M to design

277、 revolutionary materials.112 National Science Foundation.Designing Materials to Revolutionize an Engineer our Future(DMREF).(Accessed 2024 April).113 Oak Ridge National Laboratory.(2020 August).ORNL,partners receive$115 million to establish Quantum Science Center.30 transportation sector,for example

278、 enabling magnetically levitated trains that can travel with minimal friction for a smoother ride and greater efficiencytruly making science fiction come to life.Another class of materials that scientists do not know how to design are thermoelectrics,which can convert heat,even waste heat from power

279、 transmission or engines,into energy,with many applications such as cooling and electronics.Researchers have never been able to predictively design a superconductor or thermoelectric material,because these quantum materials require a unique composition of matter.114 Previous efforts have relied on c

280、ombinatorial chemical methodsempirical experimentation involving creating and screening vast numbers of material combinationswith limited success.Essentially,all discoveries of these critically important materials have been serendipitous,made by experimental trial and error.The sheer number of varia

281、bles involved,as well as the need to keep such materials affordable,makes these materials discovery problems overwhelming,and near-impossible to solve by conventional methods.There are three areas where AI tools will be a game-changer for materials science.First,the predictive abilities of AI modeli

282、ng are enabling a new approach to materials discovery by connecting and utilizing the vast quantities of data available on existing materials,their processing conditions,and their properties.From this dataset,patterns across the chemistry,physics,and engineering of materials can be determined and co

283、mbined in unique ways to provide researchers with insights and new approaches.115 Second,AI models can predict performance(for instance,predicting the coherence time of a quantum bit,the efficiency of a thermoelectric material,or the critical temperature of a superconductor),thus reducing wasteful e

284、xperimentation and testing of non-viable candidate materials.116 Third,by combining process information with material composition,practical boundaries can be placed on the material design,accelerating scale up and commercial introduction of the new materials.In addition to investigating“hard”materia

285、ls like superconductors and thermoelectrics,AI is poised to revolutionize the development of“soft”materials like polymers and fluids.Materials discovery for soft materials requires the same vast datasets and predictive capabilities as hard materials as well as the complex structure-function relation

286、ships found in material science.Unfortunately,to date,the application of AI to polymer discovery and processing remains an emerging field with vast untapped potentialmost likely the next frontier for materials science.Looking further into the future,AI tools could lead to new or improved materials s

287、uch as cold atoms,117 topological insulators,118 or superconductor-based qubits119 that serve as building blocks for quantum computers,which would be able to perform certain large-scale computations that would require impractical levels of computation or energy if performed with traditional supercom

288、puters.120 114 Keimer,B.and Moore,J.(2017 October).The physics of quantum materials.Nature Physics.115 Liu,Y.et al.(2023 December).Materials Expert-Artificial Intelligence for Materials Discovery.arXiv.116 E.g.,Biamonte,J.et al.(2017 September).Quantum machine learning.Nature.117 E.g.,Castelvecchi,D

289、.(2023 June).IBM quantum computer passes calculation milestone.Nature News.118 E.g.,Breunig,O.and Ando,Y.(2021 December).Opportunities in topological insulator devices.Nature Reviews Physics.119 E.g.,Ballon,A.(2024 March).Quantum computing with superconducting qubits.PennyLane.120 McKinsey&Company.T

290、he Rise of Quantum Computing.(Accessed 2024 April 9).31 While in recent years major progress has been made to realize small-scale quantum computing devices,these systems are unstable,susceptible to noise,and difficult to scale to the size and complexity required to compute solutions to useful proble

291、ms.Nevertheless,the potential applications of quantum computing have drawn substantial international interest and funding commitments.AI assistance could be a key factor in determining the rate of progress in this emerging technology.121 Overall,AI tools offer the opportunity,which we are starting t

292、o realize today,to usher in a new era of materials discovery and in doing so,accelerate economic growth and a new future.3.2 AI for Designing Advanced Semiconductors The modern electronic devices that underpin the global economy and our national security run on integrated circuits etched onto small

293、pieces of semiconductor material,commonly known as chips.As these chips become more capable,they also become much more complex,with current state-of-the-art chips now containing tens of billions of components.Today,only the largest corporations can afford to fabricate the most advanced chips,because

294、 of the significant engineering resources and infrastructure required to design chips.Revitalizing the U.S.semiconductor ecosystem has been a significant priority for this Administration.122 Broad usage of AI in the design of future chips can significantly increase the quality and reduce the time an

295、d number of engineers required to design the most advanced chips.For example,an aspirational goal for the U.S.semiconductor industry would be to create platforms,methodologies,and tools to enable chips to be built using a tenth of the person-hours that are required today,which would vastly lower bar

296、riers of entry to the semiconductor market,encourage innovation with a larger and more diverse set of participants,and continue to broaden our lead in semiconductor design in the world.Several AI assistants for chip designers now exist.123,124 These types of tools can allow a junior designer to ask

297、questions that previously would have consumed the time of senior designers.Some chip design AI assistants can also summarize bug reports and other design documents,or generate scripts for other design automation tools to run,all from simple English-language prompts.These AI tools do not replace desi

298、gners,but instead empower designers to be considerably more productive,helping to mitigate the overall shortage of trained chip designers.Despite circuit design being a mature field,there are AI tools available and under development that promise to provide surprising design improvements.125,126 Some

299、 of these tools are able to generate circuits that are faster or smaller than the best circuits designed using conventional methods.One 121 Hart,B.et al.(2023 August).Is China a Leader in Quantum Technologies?China Power.122 PCAST(2022 September).Revitalizing the U.S.Semiconductor Ecosystem.123 Verm

300、a,P.(2024 March).DSO.ai A Distributed System to Optimize Physical Design Flows.ISPD24 Proceedings.ISPD24:Proceedings of the 2024 International Symposium on Physical Design.124 Liu,M.et al.(2023 October),ChipNeMo:Domain-Adapted LLMs for Chip Design.arXiv.125 Roy,R.et al.(2021 December).PrefixRL:Optim

301、ization of Parallel Prefix Circuits using Deep Reinforcement Learning.2021 58th ACM/IEEE Design Automation Conference(DAC).126 Budak,A.et al.(2022 February).Reinforcement learning for electronic design automation:Case studies and perspectives.2022 27th Asia and South Pacific Design Automation Confer

302、ence(ASP-DAC),IEEE Xplore.32 mechanism for improving the performance of circuit design AI assistants is a technique known as reinforcement learning.As the AI tool explores a state-space of possible circuits receiving reinforcement via positive“reward”and negative“punishment”values for generating goo

303、d or bad circuits,respectively.It changes its approach in response to these rewards,ultimately learning which tactics lead to those circuits with desirable features.For each new semiconductor dimension,a library of thousands of cellsthe standard design building blocks of a modern chipmust be redesig

304、ned to meet the constraints of the process.For many manufacturers,this task can require about 80 person-months of effort.Using a combination of AI methods,including generative AI for clustering and reinforcement learning to fix design-rule errors,a recently developed tool127 is able to automate the

305、generation of a new library and reduce the effort by a factor of over a thousand.In a similar manner,a machine-learning based“floorplanning”tool(used to determine optimal location,shape,and size of components on a chip)uses reinforcement learning to reduce design time and improve the quality of layo

306、ut for placing these standard cells on a chip.128 In addition,field-programmable gate arrays enable fast iterations on the latest AI-driven placement and routing129 techniques demonstrating over 3x relative improvements in efficiency.130 As a chip design is created,it is subject to a number of analy

307、ses to verify whether the design meets its specifications and the constraints of the manufacturing process.AI has been applied to speed up a number of these analyses as well.For example,predicting timing before detailed routing is performed,131 or predicting undesirable“parasitic”features of a circu

308、it.132 Using AI tools such as these,a designer can iterate through many circuit ideas quickly.In the past,for instance,a layout for the circuit had to be generated to get accurate understanding of parasitic featuresoften adding days of manual effort to each iteration of the design cycle.Now the enti

309、re design iteration loop can be completed in a few minutes to obtain a circuit that meets the desired specifications.As we move forward,we expect large language models(LLMs)to evolve into design assistants that answer questions,critique and verify designs,and carry out routine design tasks.We also e

310、xpect AI techniques to significantly elevate the productivity of designers,potentially by a factor of ten or more.Designers will work at the algorithmic and system level;AI assistants will then work out the details at lower levels of design.AI synthesis and analysis tools will greatly shorten the de

311、sign cycle,allowing a design to be carried from a high-level description to verified layout in a few hours,compared to the weeks it takes today.127 Ho,C.et al.(2023 March).NVCell2:Routability-Driven Standard Cell Layout Advanced Nodes with Lattice Graph Routability Model.ISPD 23:Proceedings of the 2

312、023 International Symposium on Physical Design 128 Mirhoseini,A.et al.(2021 June).A graph placement methodology for fast chip design.Nature.129 Routing is the process of optimal path selection in any network between interconnected nodes.130 Bustany,I.et al.(2023 October).The 2023 MLCAD FPGA Macro Pl

313、acement Benchmark Design Suite and Contest Results.2023 ACM/IEEE 5th Workshop on Machine Learning for CAD(MLCAD),IEEE Xplore.131 Chhabria,V.et al.(2023 October).A Machine Learning Approach to Improving Timing Consistency between Global Route and Detailed Route.arXiv.132 Ren,H.et al.(2020 July).ParaG

314、raph:layout parasitics and device parameter prediction using graph neural networks.DAC 20:Proceedings of the 57th ACM/EDAC/IEEE Design Automation Conference.33 PCAST expects that by integrating emerging techniques like these into the chip making process,the U.S.will maintain its position as the lead

315、er in semiconductor design while relieving a critical workforce shortage in this area.3.3.Understanding and Addressing Climate Change and Extreme Weather Each year in the United States,climate disasters such as hurricanes,wildfires,and floods cause hundreds of deaths and tens of billions of dollars

316、of damage,and severely impact the physical and mental well-being of those individuals and communities forced to relocate or rebuild.While rare,the most extreme of these weather events are responsible for the majority of these costs.For instance,in 2023,the U.S.experienced damages of$92.9 billion fro

317、m 28 separate weather and climate disasters,including a drought and heat wave in the South and Midwest in the spring and fall($14.5 billion)and severe weather in the south and east in early March($6.0 billion).133 Many of these extreme events are becoming more frequent and are predicted to become ev

318、en more so with climate change,with millions of Americans projected to be at risk of being displaced from their homes by such events by the end of the century.134,135,136 In order for individuals,communities,and governments to prepare for these events and mitigate their impact,it has become increasi

319、ngly important to obtain accurate and detailed models for both weather and climate,in order to predict the trajectory of extreme weather events as they occur in real time,as well as projections of future climate risk over the longer term.Ideally,we would like to have a“crystal ball”that would show u

320、s what strength of wildfire,flood,heatwave,or hurricane one could expect to affect any given location in the next ten,twenty,or fifty years.Until recently,the best“crystal balls”available were supercomputers running huge weather and climate models that are not using AI.These older models cover the E

321、arth in a virtual grid with millions of grid points,each one simulated for millions of time steps.But these“crystal balls”are slow,blurry,and uncertain.For instance,the Global Forecast System of the U.S.National Weather Service runs four times a day in order to produce weather forecasts up to 16 day

322、s in advance,but with a spatial resolution of about 13 km,and with input data assimilated in a six hour cycle.137 For longer term climate modeling,one state-of-the-art example is the Energy Exascale Earth System Model supported by the Department of Energy,which when fully operational could produce o

323、ne possible climate outcome at a 3 km resolution for the next ten years after some weeks of supercomputer time.138 Although this is an impressive improvement,supercomputer models that do not use AI only produce a few possible outcomes among many,allowing only a limited exploration of uncertainties,a

324、nd do not pinpoint precisely where in a given grid square the extreme events are most likely to occur.133 Smith,A.(2024 January).2023:A historic year of U.S.billion-dollar weather and climate disasters.Climate.gov.134 U.S.Global Change Research Program.(2023 November).Fifth National Climate Assessme

325、nt.135 Emmanuel,K.(2017 March).Will Global Warming Make Hurricane Forecasting More Difficult?Bulletin of the American Meteorological Society.136 Bhatia,K.et al.(2019 February).Recent increases in tropical cyclone intensification rates.Nature Communications.137 National Centers for Environmental Info

326、rmation.Global Forecast System(GFS).(Accessed 2024 April 11).138 Singer,N.(2023 April).Cloud-resolving climate model meets worlds fastest supercomputer.SandiaLabNews.34 The ability of AI to directly learn from data will enhance these“crystal balls,”allowing them to become much faster and sharper.An

327、AI can be trained on data about tornadoes in Oklahoma,and it will learn how to model tornadoes in Ohio.If a scientist trains an AI on local climate and weather data,it can be used to upgrade a low-resolution climate model into a high-resolution model(for example,at 1 km resolution or less)that takes

328、 the local geography into account,a procedure known as downscaling.139 Once the AI model is trained,this process is blindingly fastthousands of times faster than traditional simulations,and thus also cheaper to run in terms of computational costs and energy usage.AI-based downscaling of coarser clim

329、ate simulations has the potential to generate hundreds or thousands of possible future outcomes of processes that cannot be resolved by the coarser models,such as tropical cyclones and storm surges,allowing a quantitative assessment of climate and weather hazards.In the future,any citizen(or a local

330、 government)will be able to enter their ZIP code into a government climate portal140 and obtain multiple detailed scenarios of expected climate changes and potential weather disasters in their area in the next ten,twenty,or fifty years.These data can then be combined with data on the built environme

331、nt to estimate the economic impact of these events.This information will be invaluable for designing appropriate building codes,emergency response preparations,insurance policies,and city planning.With regards to climate change mitigation efforts,these fast AI climate model emulators will also enabl

332、e scientists to efficiently compare the outcome of many different mitigation strategies,allowing researchers and policymakers to perform an informed cost-benefit analysis to achieve the greatest climate impact with the least cost and disruption.The speed of AI weather simulation models also makes th

333、em very suitable for real time prediction of extreme weather events,such as wildfires or hurricanes,though at the current state of the art,these models still must rely on more traditional forecasting models to provide initial data assimilation.141 For instance,a physics machine learning model,when provided with initial conditions from a traditional model,is able to provide global week-long forecas