Seminar - Differentiable Physics: A physics-constrained and data-driven paradigm for scientific discovery - Mar. 12
Romit Maulik
Assistant Professor, Data Science, Pennsylvania State University
Wednesday, Mar. 12 | 10 a.m. | AERO 114
Abstract: Machine learning stands poised to revolutionize the process of scientific discovery across various disciplines. In this talk, we will introduce a state-of-the-art scientific machine learning paradigm - differentiable physics (DiffPhys). DiffPhys can be considered a system identification paradigm that can be applied to determine neural network approximations of governing laws given data. It can also be used to improve first-principles-based simulations of physical phenomena by learning corrections to governing laws (for instance for closure modeling in multiscale applications). Notably, optimizing these neural networks necessitates a differentiable programming paradigm where gradients of a loss function can be propagated through a numerical solver. In this talk, we will introduce DiffPhys algorithms that (1) can learn models for dynamical systems from sparse data, (2) efficiently compute sensitivities for systems exhibiting deterministic chaos, (3) leverage graph neural networks for geometry-invariant learning, and (4) provide physically meaningful interpretations for neural network behavior thereby engendering scientific discovery. We will demonstrate the capabilities of DiffPhys on canonical and realistic scientific computing problems and close with a discussion of the future possibilities of this approach.
Bio: Romit Maulik is an Assistant Professor of Data Science in the College of Information Sciences and Technology at Pennsylvania State University and a Joint Appointment Faculty at the Mathematics and Computer Sciences Division at Argonne National Laboratory (Argonne). He obtained his PhD in Mechanical and Aerospace Engineering at Oklahoma State University in 2019 and was the Margaret Butler Fellow and then a Staff Scientist at Argonne National Laboratory before joining Penn State in 2023. His research centers around machine learning for scientific computing with an emphasis on scalable, physically consistent, and robust algorithm construction for simulation-based scientific discovery of multiscale physics from multifidelity data. He has led research projects sponsored by multiple agencies and is an Early Career Awardee of the Army Research Office.