Interpretable and Structure-Preserving Data-Driven Methods for Physical Simulations

Explore a comprehensive presentation on interpretable and structure-preserving data-driven methods for physical simulations. Delivered by Youngsoo Choi from Lawrence Livermore National Laboratory, this 51-minute talk delves into the importance of physical simulations in modern science and examines various data-driven approaches. Learn about conditional generative adversarial networks, the pros and cons of black-box approaches, and methods to achieve interpretability. Discover how Dynamic Mode Decomposition (DMD) accelerates 3D printing process simulations and how time-windowing Wavelet DMD improves accuracy. Investigate other interpretable data-driven methods, including Parameterized Latent Space Dynamics Identification (LaSDI) and its application to radial advection problems. Explore physics-constrained models, projection-based linear subspace reduced order models, and space-time ROM for maximal compression. Gain insights into component-wise ROM for lattice-structure design optimization, PROM for wind turbine blade design optimization, and database local ROMs for multi-start airplane wing optimization. Conclude with an overview of data-driven methods categorized by their level of intrusiveness.