On Tuesday, 25 Feb at 2:00 p.m. in NSII 1201 Prof. Jack Xin will present a seminar for all Physical Sciences researchers interested in machine learning (title and abstract below). This seminar series is offered every other Tuesday at 2 in NSII 1201 to inform, engage, and facilitate artificial intelligence research activities throughout all four PS departments and associated research collaborations.
Attend activities of the Machine Learning Nexus in Physical Sciences as useful for your research and career. Contribute to, and keep track of, the Machine Learning Nexus at https://ps.uci.edu/psml/.
Seminar Information
Speaker: Prof. Jack Xin, Math
Location: 1201 NSII
Date and Time: Tuesday, 25 February at 2:00 p.m.
Title: Finding Low Dimensional Representation of Multi-scale Dynamics via Deep Adaptive Basis Learning
Abstract:
Low dimensional representation of multi-scale dynamics allows fast computation of complex physical problems. In fluid turbulence for example, small scales form dynamically and render fully resolved simulations expensive. Methods using Fourier or other analytical representation require exceedingly large number of basis functions in the small viscosity regime. The classical data-driven approach to adaptive basis learning dated back to 1967 is known as proper orthogonal decomposition (POD) which performs singular value decomposition (a.k.a. principal component analysis) on observations (snapshots of computed solutions) at a sample viscosity value. However, POD basis can lose representation capability quickly when applied to smaller viscosity values. I shall show a deep learning framework using super-resolution neural networks to model the transition of physical solutions and improve POD predictability. An application is computing large scale diffusivity of Brownian particle in a chaotic flow.