Joint Seminar: Learning Reduced Order Models for Fluid Mechanics with Interpretable Machine Learning
Turbulence in fluids is a highly non-linear, chaotic phenomenon with vast applications in applied science and engineering. However, the intense multi-scale nature of turbulence greatly limits the scales at which numerical solutions can be resolved given computational constraints. Therefore, developing reduced-order models (ROMs) for such applications has been an area of intense research, especially with the superior representational capacity of Deep neural networks. However, these strategies are typically devoid of physical foundations, constraints and often lack interpretability. In this talk, I present two strategies we are developing to address these limitations by *explicitly* incorporating physics into neural networks as hard constraints: a) A differentiable programming based ROM approach that incorporates neural networks into the projected Navier-Stokes equations in Galerkin form, and b) Guaranteeing inductive bias for specific physics constraints by embedding them into traditional "black-box" neural networks such as CNNs and GANs. The talk will outline our results from both these strategies, which demonstrate improved accuracy, reduced computational costs and superior interpretability, and present future directions.
Date
08.03.2023
Time
13:30 h
Place
- Bundesstr. 53, room 022/023
- Seminar Room 022/023, Ground Floor, Bundesstrasse 53, 20146 Hamburg, Hamburg
Speaker
Arvind T. Mohan, Los Alamos National Laboratory
Dr. Arvind Mohan is a scientist in Computational Physics and Methods group at Los Alamos National Laboratory. He obtained his Ph.D. in Aeronautical and Astronautical Engineering from The Ohio State University. His expertise is in Fluid dynamics, machine learning and reduced-order modeling of high-dimensional complex systems. His current research is focused on developing interpretable deep learning algorithms for surrogate modeling of PDEs. His other research efforts span diverse areas at LANL, including nuclear physics, earth sciences, and astrophysics, with the focus being on scientific machine learning.