Talk starts at 1:45
Prof. Nicholas Zabaras speaking in the UW Data-driven methods in science and engineering seminar on May 14, 2021.
Affiliation: Scientific Computing and Artificial Intelligence (SCAI) Laboratory, University of Notre Dame
Website: www.zabaras.com/
For more information including past and upcoming talks, visit: www.databookuw....
Sign up for notifications of future talks: mailman11.u.wa...
Abstract: Most attempts at the interface of physical modeling and machine learning (ML) employ computationally-generated data in order to drive the various statistical discovery objectives. This enables a direct transfer and application of ML tools and techniques, but removes valuable structure, symmetries and invariances that were present in the model. In order to rediscover this structure, if at all possible, ML tools would need copious amounts of data. Even when big data is available, it is important to ensure that predictions produced by ML models trained on this data satisfy these constraints. Purely data-based, modern ML tools as those based on deep neural networks (DNN) provide rich representations for learning complex nonlinear functions, but lack robustness and fail when higher-level abstractions implied by the physical structure are needed to make predictions. We will discuss some aspects in the development of generative multiscale deep learning approaches incorporating known physical constraints as prior knowledge. Examples will be provided in the development of DNN surrogate models for multiscale stochastic PDEs, multi-fidelity DNN algorithms for fluid flows, integrating pre-trained VAE models with high-dimensional Bayesian multiscale inversion tasks, Koopman embedding in transformer models for prediction of dynamical systems, and deep generative modeling of molecular dynamics.
13 май 2021
