Featured Speakers: Per-Olof Persson, Silvia Crivelli, Ben Erichson
Date: July 1, 2025
Time: 11 a.m. – 12 p.m.
Where: 59-3101. Virtual: Zoom Available
Three Short Talks:
High-Order Numerical Methods for Simulations in Science and Engineering
Presented by Per-Olof Persson
(Mathematics Group, AMCR and Professor of Mathematics, UC Berkeley)
Numerical simulations play a central role in advancing both engineering and scientific research. By combining mathematical modeling with computational power, they enable us to predict complex physical behavior and gain insights that drive innovation in design, optimization, and analysis. In particular, models based on partial differential equations (PDEs) are fundamental to a wide range of applications across the applied sciences. In this talk, I will give a broad overview of recent advances in high-order accurate finite element methods, which offer improved precision and efficiency in solving such models. I will also introduce the important topic of unstructured mesh generation, highlighting both my DistMesh software and a new approach based on deep reinforcement learning.
AI x Medicine: How AI Can Heal Healthcare
Presented by Silvia Crivelli
(Program Manager, Computational Biosciences, SciData and Associate Researcher, UC Davis)
Despite spending more on healthcare than any other nation, the United States lags behind other developed countries in life expectancy—largely due to a system built around single-disease treatment and “one-size-fits-all” care. In reality, patients often face complex chronic conditions, multimorbidity, and highly individualized responses to treatment. Large Language Models (LLMs) offer a transformative opportunity to shift this paradigm. Clinical LLMs—developed across both academia and industry—show immense promise for improving care delivery and enabling precision medicine. However, their progress is limited by a critical bottleneck: access to comprehensive, high-quality clinical data. In an unprecedented collaboration, the U.S. Department of Energy (DOE) and the Department of Veterans Affairs (VA) are addressing this challenge head-on. With access to genomic and electronic health record (EHR) data from millions of Veterans, teams across eight DOE national laboratories are building AI models to tackle high-impact conditions including mental health disorders, obstructive sleep apnea, long COVID, and lung cancer. In this talk, I will present the development of a clinical LLM trained on this vast, multimodal, and longitudinal dataset. I’ll highlight how this work not only advances AI in healthcare, but also lays the foundation for a more personalized, effective, and equitable healthcare system.
Modeling Spatio-temporal Physical Systems with Diffusion Models
Presented by Ben Erichson (Machine Learning & Analytics Group, SciData)
Modeling physical systems that change over space and time is often done using partial differential equations (PDEs). However, these mathematical models usually rely on simplifying assumptions, which can make it hard to describe complex real-world systems, especially when the underlying physics is not fully known. Generative AI models offer a different, data-driven way to approach this problem. They don’t need strong assumptions and can scale to large datasets.In this talk, I will give a short introduction to score-based diffusion models, which generate realistic outputs by gradually turning random noise into structured data. I’ll show how these models can be used for tasks like super-resolution and forecasting in fields such as climate and fluid dynamics. I’ll also explain how repeating the generation process can give useful estimates of uncertainty, which is important when working with physical systems. Overall, diffusion models are a promising tool for learning from complex spatio-temporal data.
Read more about our presenters below.
Featured Speakers: Katie Klymko, Bert de Jong
Date: July 3, 2025
Time: 11 a.m. – 12 p.m.
Where: 59-4102. Virtual: Zoom Available
Berkeley Lab scientists describe their path from education to a career at Berkeley Lab. Insightful and inspirational, this inside look proves there’s no single path to great scientific research.
Autobiographical insights provided by Katie Klymko (Advanced Technologies Group, NERSC) and Bert de Jong (Department Head, Computational Sciences and Group Lead of Applied Computing for Scientific Discovery, AMCR).
Per-Olof Persson is a Professor of Mathematics at the University of California, Berkeley, and a Faculty Scientist at Lawrence Berkeley National Laboratory. His research focuses on high-order discontinuous Galerkin methods for computational fluid and solid mechanics, including efficient discretizations, scalable solvers, and adjoint-based optimization. He has also contributed extensively to mesh generation, developing methods for space-time and curved meshes, as well as new approaches based on machine learning.
Dr. Crivelli has conducted research at the intersection of science, high-performance computing, human-computer interaction, and applied mathematics for 30 years. Her research has focused on two main goals: 1) to bring scientists together, both seasoned and young and from all walks of science, to tackle long-standing, extremely hard, and multidisciplinary problems and 2) to develop methods and software tools that empower physicians and researchers to predict the behavior of biological systems and, more recently, healthcare outcomes. See her work at the Crivelli Group website: https://crivelligroup.lbl.gov/
Ben Erichson is a Research Scientist at Lawrence Berkeley National Laboratory and leads the Robust Deep Learning Group at the International Computer Science Institute (ICSI), an affiliated institute of UC Berkeley. Prior to this role, he was a Tenure-Track Assistant Professor of Data-driven Modeling and Artificial Intelligence at the University of Pittsburgh. His research centers on the intersection of deep learning and dynamical systems, with the goal of developing robust neural network architectures for scientific applications. Currently, his interests include generative AI for spatio-temporal modeling and the development of foundation models for science.
Katie Klymko received her PhD in 2018 from UC Berkeley in theoretical chemistry. She was a postdoc at LBL from October of 2018 through September of 2021 focused on quantum algorithms for eigenvalue calculations in molecular systems as well as algorithms to explore thermodynamic properties. In October of 2021, she became a staff member at NERSC where she is working to integrate quantum into future HPC workloads.
Bert de Jong serves as the Department Head for Computational Sciences, and leads the Applied Computing for Scientific Discovery Group, which advances scientific computing by developing and enhancing applications in key disciplines, as well as developing tools and libraries for addressing general problems in computational science.
