CSE Community Seminar | October 2, 2026

Abstract

Modern foundation models are remarkably effective at learning complex distributions from data, but scientific applications impose a stricter requirement: outputs must not only be plausible, but also satisfy governing equations, conservation laws, boundary conditions, and other physical constraints. This talk explores how such constraints can be incorporated into scientific ML without requiring a specialized architecture or retraining a model from scratch for every new physical problem.

I will discuss methods for introducing physical constraints at different stages of the learning pipeline, including probabilistic modeling, post-training, and test-time steering methods. These ideas span different classes of foundation models, from generative image models for scientific fields to language models that generate scientific code. A recurring theme is how physical constraints can move from being fixed architectural choices to reusable computational objects that can guide, correct, or evaluate models when needed.

More broadly, the talk will explore what it means to build computational foundations for scientific ML in which learning, scientific structure, and scalable computation are designed together, with a particular focus on PDE-governed systems and scientific simulation.