DD-FEM: A Physics-Governed Path to Foundation Models | Youngsoo Choi | JHU-IITD SMaRT скачать в хорошем качестве

DD-FEM: A Physics-Governed Path to Foundation Models | Youngsoo Choi | JHU-IITD SMaRT

This talk is part of the Scientific Machine Learning Research Talks (SMaRT) Seminar Series, a joint initiative between Johns Hopkins University and IIT Delhi. 🔹 Speaker: Youngsoo Choi, Staff Scientist, Lawrence Livermore National Laboratory, USA 🔹 Talk Title: DD-FEM: A Physics-Governed Path to Foundation Models for Computational Science 🔹 Date: Wednesday, February 11, 2026 Time: 8:30 PM IST | 3:00 PM GMT | 10:00 AM ET 📄 Abstract Scientific AI increasingly needs models that can be reused, adapted, and trusted across changing geometries and regimes. Yet many neural surrogates remain interpolation engines tied to a narrow training distribution. This talk argues that a scientific foundation model must provide not only predictive power, but also structural invariances and auditable, governed composition. I present a definition and evidence-based readiness criteria for foundation models in computational science, then describe DD-FEM, a framework that modernizes finite elements by replacing hand-designed polynomial bases with data-driven local bases, while preserving the FEM principle of local computation + global assembly. DD-FEM supports multiple assembly/coupling mechanisms (including static condensation and DG) and enables modular transfer: learned local components can be reused across larger domains, new configurations, and—in some cases—new PDE families. Across representative PDEs (elasticity, incompressible flow, Burgers, Poisson), DD-FEM delivers order-of-magnitude speedups in forward simulation and repeated-query settings such as inverse problems and design optimization, while maintaining percent-level fidelity. I conclude with open problems and a roadmap for building scientific foundation models that generalize, assemble, and endure. 👤 About the Speaker Youngsoo Choi is a staff scientist in LLNL’s Center for Applied Scientific Computing (CASC), where he develops reusable, scalable learning frameworks for accelerating scientific simulation. His work spans reduced-order modeling, data-driven surrogates, inverse problems, uncertainty quantification, and optimization, and includes methods such as nonlinear-manifold ROMs, space–time ROMs, and latent-dynamics identification. He leads the libROM team and contributes to open-source software including libROM and pylibROM. He received his B.S. from Cornell and Ph.D. from Stanford, and held postdoctoral positions at Sandia National Laboratories and Stanford University before joining LLNL in 2017. 📌 Follow the SMaRT Seminar Series for more talks at the intersection of AI, physics, and scientific computing. #ScientificMachineLearning #StructuralHealthMonitoring #PhysicsEnhancedML #UncertaintyQuantification #BayesianInference #SHM #AppliedMechanics #PhysicsInformedAI #JHUIITD #SMaRTSeries