Published on 2024-09-13 by qzhe
I was pleased to share our recent research on Differentiable Solid Mechanics and Geophysics with students and colleagues during my visit to Singapore in September. The talk was titled "Neural-Integrated and Data-Driven Approaches for Computational Modeling in Solid Mechanics and Geophysics." I am grateful for the warm invitations from NUS, NTU and A*STAR.
- ME Department Seminar, National University of Singapore (NUS)
- iHPC Seminar, Agency for Science, Technology and Research (A*STAR)
- CEE Department Seminar, Nanyang Technological University (NTU)
Title: Neural-Integrated and Data-Driven Approaches for Computational Modeling in Solid Mechanics and Geophysics
Abstract: We present the development of a hybrid computational framework that integrates physics-based numerical schemes with machine learning methods to address various forward and inverse problems in computational mechanics. Our focus is on applications involving inherent material complexities and coupling effects, and exploring how underlying physics laws can be effectively imposed in these methods when different amounts of data are available. We begin by introducing a physics-informed machine learning approach, termed Neural-Integrated Meshfree (NIM), designed to overcome the issues of low accuracy and training efficiency in simulating large-deformations and material nonlinearities. This approach utilizes a hybrid approximation strategy that combines neural network representations with customized basis functions. The effectiveness of the NIM method is demonstrated through various linear and nonlinear benchmark mechanics problems, including identification of heterogeneous biological materials. Additionally, in scenarios where experimental/simulation data are abundant, we introduce a hybrid scheme that leverages data-driven learning models for solving different coupled systems. We show that the proposed machine learning models can reliably learn operators that capture the underlying physical processes, enabling reduced-order modeling of complex, nonlinear high-dimensional problems. Geophysical and biological examples will be presented to showcase the versatility of these machine learning techniques in enhancing scientific computing.
