Science in Parallel

Foundation models-- LLMs or LLM-like tools-- are a compelling idea for advancing scientific discovery and democratizing computational science. But there's a big gap between these lofty ideas and the trustworthiness of current models.

Youngsoo Choi of Lawrence Livermore National Laboratory and his colleagues are thinking about to how to close this chasm. They're engaging with questions such as: What are the essential characteristics that define a foundation model? And how do we make sure that scientists can rely on their results?

In this conversation we discuss a position paper that Youngsoo and his colleagues wrote to outline these questions and propose starting points for consensus-based answers and the challenges in building foundation models that are robust, reliable and generalizable. That paper also describes the Data-Driven Finite Element Method, or DD-FEM, a tool that they've developed for combining the power of AI and large datasets with physics-based simulation.

You'll meet:

Youngsoo Choi is a staff scientist at Lawrence Livermore National Laboratory (LLNL) and a member of the lab's Center for Applied Scientific Computing (CASC), which focuses on computational science research for national security problems. Youngsoo completed his Ph.D. in computational and mathematical engineering at Stanford University and carried out postdoctoral research at Stanford and Sandia National Laboratories before joining Livermore in 2017.

Computational scientists can take on the role of utility players in research, and Steven Wilson is one example. At Arizona State University he's built instruments, carried out experiments and dove deep into computational work. As a postdoc, he's working on a new challenge: building a quantum chemistry startup company. In this episode, he discusses his career that started with 10 years in the United States Navy Nuclear Program, how that military experience shaped his academic studies and the role of the Department of Energy Computational Science Graduate Fellowship (DOE CSGF) in shaping his research to make chemical reactions more efficient.

You'll meet:

• Steven Wilson is a postdoctoral researcher in Christopher Muhich's research group at Arizona State University, where he completed both his undergraduate degree in 2020 and his Ph.D. in 2024. He was a DOE CSGF recipient from 2021 to 2024 and completed practicum research at Pacific Northwest National Laboratory (PNNL). He is also CEO of PsaiForge, a quantum chemistry software startup that he cofounded with Muhich.

While we take a short summer break, we're posting one of our favorite past episodes and a great follow-up to our last episode with Amanda Randles of Duke University. In 2023, we talked with Joe Insley of Argonne Leadership Computing Facility and Northern Illinois University about data visualization, from the practical process of helping researchers understand their results to showstopping images and animations that make the work accessible to broad audiences. Joe discusses his career path, how he and his team approach visualization projects, his work with students and his advice for improving scientific data visualization.

You'll meet:

• Joe Insley is team lead for visualization and data analysis at Argonne Leadership Computing Facility and associate professor in the School of Art and Design at Northern Illinois University. Joe got his start in scientific visualization creating interactive data explorations for the CAVE (cave automatic virtual environment).

Duke University associate professor Amanda Randles' work to simulate and understand human blood flow and its implications demonstrates how high-performance computing paired with scientific principles can help improve human health. In this conversation, she talks about how she brought together early interests in physics, coding, biomedicine and even political science and policy and followed her enthusiasm for the Human Genome Project. She discusses how supercomputers are pushing the boundaries of what researchers can learn about the circulatory system noninvasively and how that knowledge, paired with data from wearable devices, could lead to new ways to monitor and treat patients. She also talks about her public engagement and science policy work and its importance, both for educating patients and supporting computational science's future.

You'll meet:

Amanda Randles is the Alfred Winborne and Victoria Stover Mordecai associate professor of biomedical sciences at Duke University and director of Duke's Center for Computational and Digital Health Innovation. Her research using high-performance computing to model the fluid dynamics of blood flow has garnered numerous awards including one of the inaugural Sony Women in Technology Awards with Nature , the 2024 ISC Jack Dongarra Early Career Award and the 2023 ACM Prize in Computing. Amanda completed her Ph.D. at Harvard University working with Efthimios Kaxiras and Hanspeter Pfister. She was a Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient from 2010 to 2013 and a Lawrence Fellow at Lawrence Livermore National Laboratory from 2013 to 2015.

Follow Amanda on social media: LinkedIn, BlueSky and Instagram.

Understanding how the brain works remains a grand scientific challenge, and it's yet another area where researchers are examining whether foundation models could help them find patterns in complex data. Joel Ye of Carnegie Mellon University talks about his work on foundation models, their potential and limitations and how others can get involved in applying these AI tools.

Joel Ye is a Ph.D. student in the program in neural computation at Carnegie Mellon University in Pittsburgh, where he studies ways to understand brain data and brain-computer interfaces. He's a third-year Department of Energy Computational Science Graduate Fellow.

Chemists and chemical engineers have modeled molecules for decades, but artificial intelligence and foundation models offer the prospect that researchers could train models with predictive abilities in one area of chemistry that could be fine-tuned for another. Trustworthy chemistry foundation models could help streamline the experimental time and resources needed to discover new medicines or design new batteries. Massachusetts Institute of Technology Ph.D. student Jackson Burns is working on these questions. He describes the promise and challenges of building foundation models in chemistry, his work on chemprop, and his advice to other researchers interested in working on foundation models for chemistry and science in general.

You'll meet:

Jackson Burns is a Ph.D. student in William Green's chemical engineering group at MIT. He's also a third-year Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient. He completed his undergraduate degree in chemical engineering at the University of Delaware.

In the second episode in our series on foundation models for science, we discuss Oak Ridge National Laboratory's work and hear about lessons learned from the recent 1000 Scientists AI Jam, a recent event that brought together researchers from several Department of Energy national laboratories, OpenAI and Anthropic. My guest is Prasanna Balaprakash, ORNL's director of AI programs. We talk about how foundation models could help with climate forecasts and his team's 2024 Gordon Bell finalist research and futuristic work that applies principles of swarm intelligence for managing distributed computing resources.

Prasanna Balaprakash has been the director of artificial intelligence programs at Oak Ridge National Laboratory (ORNL) since March 2023. Previously he had worked as a postdoctoral researcher and staff computer scientist at Argonne National Laboratory. He was a 2018 recipient of a Department of Energy Early Career Research Program award.

Large language models aren't just powering chatbots like ChatGPT. This type of computational model is an example of a particular flavor of artificial intelligence known as foundation models, which are trained on vast amounts of data to make inferences in new areas. Although text is one rich data source, science offers many more from biology, chemistry, physics and more. Such models open up a tantalizing new set of research questions. How effective are foundation models for science? How could they be improved? Could they help researchers work on challenging questions? And what might they mean for the future of science?

This episode begins a series where we'll explore these questions and more, talking with computational scientists about their work with foundation models and the opportunities and challenges in this exciting, rapidly changing area of research. We'll start by talking with Ian Foster of Argonne National Laboratory and the University of Chicago about AuroraGPT, a foundation model being developed for science and named for Argonne's new exascale computer.

You'll meet:

Ian Foster is a senior scientist at Argonne National Laboratory where he directs the data science and learning division. He's also a professor of computer science at the University of Chicago. He is the co-leader of the data team for Argonne's AuroraGPT project.

Wrapping up our discussion of the 2024 Nobel Prizes in Physics and Chemistry, computer scientist Mansi Sakarvadia and computational structural biologist Josh Vermaas talk about the recent prizes and what they mean for science. You'll hear about how the prizes both break down research barriers and introduce concerns about misinformation and public trust. The research honored with the chemistry prize has already changed how researchers study questions that involve understanding proteins' structures. For more on the 2024 Nobel Prizes, check out our recent interview with Anil Ananthaswamy.

You'll meet:

Mansi Sakarvadia is a Ph.D. student in the computer science department at the University of Chicago and a current Department of Energy Computational Science Graduate Fellow. She studies ways to interpret how machine learning models work.

Josh Vermaas is an assistant professor at Michigan State University. His research in computational structural biology focuses on understanding photosynthesis and energy transfer processes in plants as part of the MSU-DOE Plant Research Laboratory.

2024 was artificial intelligence's Nobel Prize year with the physics and chemistry prizes recognizing the underpinnings and application of these algorithms. Science journalist and author Anil Ananthaswamy spent years writing a popular book, Why Machines Learn: The Elegant Math Behind Modern AI, that explores the equations and historical context for this technology.

In this conversation, Anil and host Sarah Webb explore that math and history, the significance of these Nobel Prizes for both AI and science, and the challenges that come with this powerful and fast-moving technology.

You'll meet:

Anil Ananthaswamy is an award-winning journalist and journalist-in-residence at the Simons Institute for the Theory of Computing at the University of California, Berkeley. Previously he has worked as a staff writer and editor for New Scientist magazine. He has written four books including Why Machines Learn: The Elegant Math Behind Modern AI (Dutton, 2024).