Abstract
Physics-constrained data-driven computing is an emerging paradigm that directly integrates material database into physical simulations of complex materials, bypassing the construction of classical constitutive models. However, most of the developed data-driven computing approaches are based on simplistic distance minimization and thus suffer from dealing with high-dimensional applications and lack generalization ability. This study proposes a deep learning enhanced data-driven computing framework to address these fundamental issues for nonlinear materials modeling. To this end, an autoencoder, a special multi-layer neural network architecture, is introduced to learn the underlying low-dimensional embedding representation of the material database. Incorporating the offline trained autoencoder and the discovered embedding space in online data-driven computation enables to search for the optimal material state from database in low-dimensional embedding space, enhancing the robustness and predictability of data-driven computing on limited material data. To enhance stability and convergence of data-driven computing, a convexity-preserving interpolation scheme is introduced for constructing the material state on the low-dimensional embedding space given by autoencoders. The effectiveness and enhanced generalization performance of the proposed approach are examined by modeling biological tissue with experimental data.
Original language | English (US) |
---|---|
Article number | 67 |
Journal | CEUR Workshop Proceedings |
Volume | 2964 |
State | Published - 2021 |
Externally published | Yes |
Event | AAAI 2021 Spring Symposium on Combining Artificial Intelligence and Machine Learning with Physical Sciences, AAAI-MLPS 2021 - Stanford, United States Duration: Mar 22 2021 → Mar 24 2021 |
Bibliographical note
Publisher Copyright:Copyright © 2021for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0)