1

Introduction

This chapter presents the rationale for the study, including any directive that led to its initiation. The statement of task for the study, the committee’s interpretation of elements of the statement of task, and the structure of the report are also described. The chapter also points to potential reasons why the Department of Energy (DOE) is furthering the development and use of artificial intelligence (AI) models, such as foundation models. The study charge and the committee’s interpretation of its key elements are then discussed, followed by a review of the report’s structure in fulfillment of the study charge.

SIGNIFICANCE OF FOUNDATION MODELS

Foundation models are typically large-scale neural networks trained on vast amounts of heterogeneous data with the capability of learning new representations via fine-tuning on additional data. They represent a departure from traditional AI systems designed for specific tasks. They can be standalone systems or can be used as a “base” for many other applications (see Figure 1-1). Today, the most prominent foundation models are large language models (LLMs) trained on vast amounts of text data to process and generate human-like responses, answer follow-up questions, and complete other language-related tasks. There is widespread enthusiasm about the use of foundation models, especially LLMs and approaches that build on LLMs, to advance scientific research (Lee 2024).

When these models are used in scientific research, they encounter challenges including limited domain-specific knowledge, interpretability of the results, sparse training data, integration with experimental data, lack of causal understanding, and the evolving nature of scientific knowledge.

These challenges provide opportunities for research across all areas of DOE. The pursuit of these opportunities is an important endeavor as the private sector is presently leading the race for the development of state-of-the-art foundation models. The landscape of this race is in constant flux, and the leaders at any time will reap major rewards and may determine the direction of future scientific endeavors.

The DOE national laboratories are special-purpose entities referred to as federally funded research and development centers (FFRDCs). FFRDCs provide the government with a dedicated, objective, and highly specialized technical and analytical capability that is essential for addressing long-term, complex national challenges. FFRDCs cannot manufacture products or compete directly with industry and have no commercial or shareholder interests, ensuring that their advice, analysis, and research are unbiased, allowing them to act as “honest brokers” and trusted advisors. They attract, develop, and retain unique scientific expertise that combines world-class research and entrepreneurial know-how to support the mission of the agencies they serve. By assembling teams of experts from various fields, FFRDCs address multifaceted technical challenges that often require high-risk experiments and large facilities, such as supercomputers or light sources. FFRDCs play a crucial role in maintaining and advancing the nation’s

scientific and technical expertise in critical areas and facilitate technology transfer to the private sector. As such, the DOE national laboratories have an important role to play in advancing AI technologies, particularly AI foundation models for scientific discovery and innovation.

AI, particularly with the emergence of foundation models, is a transformative force poised to redefine future economies, national security, scientific discovery, global power dynamics, and daily life. Given this immense impact, maintaining U.S. leadership in AI is imperative, necessitating an understanding of the global competitive landscape, particularly coming from China.

China has strategically prioritized its development of AI, aiming to become a world leader in the field by 2030. This goal was outlined in its “Next-Generation Artificial Intelligence Development Plan,” which was released in July 2017. Their ambition is supported by significant government investment in AI theory, technology, and application. Chinese AI firms have expanded their influence by freely distributing their models for the public to use, download, and modify, which makes them more accessible to researchers and developers around the world. In terms of quantifiable metrics, China is ahead of the United States: it significantly outpaced the United States in AI patent filings in 2022, possesses a leading advantage in the sheer volume of data, and leads the United States in the quantity of AI scientific papers. China has cultivated a robust domestic ecosystem, boasting abundant science, technology, engineering, and mathematics talent, resilient supply chains, and impressive manufacturing capabilities (Omaar 2024).

The nation that shapes the LLMs powering tomorrow’s applications and services will wield great influence not only over the norms and values embedded in them but also over the critical semiconductor ecosystem that underpins AI computing. The fact that both China and the United States believe that these technologies could also provide military advantages only heightens the importance of achieving and maintaining long-term AI leadership.

Although the report will be examining use of foundation models for scientific discovery and innovation specifically for DOE, the development and use of these tools will benefit the general scientific community. The report will examine how foundation models can help drive progress in complex systems—such as digital twins—and unlocks new findings in areas vital to American competitiveness, including materials science, nuclear science, and public health.

STUDY APPROACH

The study was supported by DOE’s Office of Science, National Nuclear Security Administration, and Biological and Environmental Research program. In collaboration with the National Academies of Sciences, Engineering, and Medicine, these DOE offices developed the study’s statement of task (see Box 1-1). The National Academies appointed a committee of 11 members with expertise in mathematics, statistics, computer science, data science, algorithms and scal-

ability, energy consumption and computing, scientific applications, model trustworthiness, and DOE and laboratory experience. Committee biographies are provided in Appendix D.

The committee held several information-gathering meetings in support of this study, including one in-person public meeting (March 11–12, 2025) where the committee was presented with material from industry scientists and AI leaders from DOE laboratories. The other information-gathering sessions (February 11, May 6, and May 20, 2025) were virtual where presenters discussed DOE’s interest in AI for science, learning models from data, and agentic AI.

Report Organization

This report was written with the intention of informing the scientific and research community, academia, pertinent government agencies, AI practitioners, and those in relevant industries about open needs when developing and using foundation models. The study takes an objective approach to understanding the field of foundation models specifically for scientific discovery and innovation and the potential opportunities that their use and development can bring to DOE. The report begins with a discussion on the use of foundation models with and without traditional modeling techniques¹ (Chapter 2). Chapter 3 explores the suc-

___________________

¹ For this report, traditional modeling refers to large-scale computational science solvers as well as statistical models.

cesses and exemplar use cases of foundation models and potential applications in which DOE could be most successful in its endeavors with foundation models for science. Chapter 4 discusses the strategic considerations and directions of foundation model use while challenges that the use of foundation models impose are covered in Chapter 5. The committee addresses major conclusions and recommendations throughout Chapters 2 through 5.

The committee would like to stress that while the report uses the terms AI, AI for science, AI models, LLMs, machine learning, and foundation models, the report is specifically directed toward the use and development of foundation models for science. The report is further specifically directed toward DOE’s use and development of these models.

REFERENCES