Wed, Mar 09, 2022 @ 03:30 PM - 04:30 PM
Aerospace and Mechanical Engineering
Conferences, Lectures, & Seminars
Talk Title: Data-driven discovery of governing equations with deep learning and sparse identification techniques
Abstract: Machine learning techniques promise to offer the ultimate form of automation, particularly when applied to computational modeling and simulation. As a consequence, the computational scientist\'s narrative now revolves around discovering physics directly from data, with as little assumptions about the underlying physical system as possible. I briefly go over the latest attempts to accomplish this goal and focus on my recent work in combining deep learning with sparse identification of differential equations. First, I show how probability distribution function (PDF) equations can be inferred from Monte Carlo simulations for coarse-graining and closure approximations. Second, I present our latest results on discovering dimensionless groups from data, using the Buckingham Pi theorem as a constraint. And third, I go over the deep delay autoencoder algorithm that reconstructs high dimensional models from partial measurements as motivated by Takens\' embedding theorem. I finally highlight the limitations of these methods and propose a few directions for future research.
Biography: Joseph Bakarji is currently a postdoctoral fellow in the department of mechanical engineering at the University of Washington, working with Steven Brunton and Nathan Kutz. He received his PhD in 2020 from Stanford University where he developed multiscale stochastic models for granular materials and data-driven closure models for uncertainty quantification. Joseph received the Henry J. Ramey, Jr. and the Frank G. Miller fellowship awards in 2018 and 2020 respectively. His current research focuses on combining deep learning and sparse identification methods, to discover interpretable physical models in complex systems from data.
More Info: https://usc.zoom.us/j/93987337017?pwd=MWd2dXBSL1FaR1RPaHNscjJ1NW80UT09
Audiences: Everyone Is Invited
Contact: Tessa Yao