From Foundation Models to Data-Driven Finite Elements

Content

Featured Speaker:

Youngsoo Choi
Staff Scientist in the Center for Applied Scientific Computing at Lawrence Livermore National Laboratory

From Foundation Models to Data-Driven Finite Elements

October 22, 2026
2 PM 
Sustainable Engineering Laboratories, Room 101 

Part of the AI Engineering Seminar Series
 

Abstract:

Foundation models have transformed language modeling by providing reusable and adaptable models for many downstream tasks. Their success raises a central question for computational science: what should a foundation model mean when simulations span different physical systems, geometries, parameter regimes, and scales? I will propose a definition inspired by both modern machine learning and foundational numerical methods such as finite elements. Scientific foundation models should be general, reusable, scalable, and physically consistent, respecting governing equations, conservation laws, interface conditions, and other mathematical structure required for trustworthy simulation. Using these criteria, I will assess emerging approaches and introduce the data-driven finite element method (DD-FEM) as a potential realization of this vision. DD-FEM learns reusable local representations from small training problems and assembles them to solve new and much larger systems. I will discuss several basis-construction and assembly strategies and demonstrate the framework in structural design, porous-media flow, time-dependent Burgers and incompressible Navier–Stokes equations, and plasma transport.

Short Bio: 

Youngsoo Choi is a Staff Scientist in the Center for Applied Scientific Computing at Lawrence Livermore National Laboratory, where he leads research in scientific machine learning, reduced-order modeling, and data-driven physical simulation. He leads the libROM team, serves as LLNL PI for the DOE SciDAC LEADS Institute, and currently supports approximately a dozen research projects and programs spanning DOE, national-security, two GENESIS projects, and LLNL initiatives. His research has developed a broad family of latent-space dynamics identification methods, including LaSDI and its extensions, as well as scalable domain-decomposition and data-driven finite element approaches aimed at reusable, physics-consistent foundation models for computational science. His work spans fluid dynamics, plasma physics, hydrodynamics, additive manufacturing, first-principles molecular dynamics, geophysics, and scientific optimization. Dr. Choi received his Ph.D. in Computational and Mathematical Engineering from Stanford University.