| About this Abstract |
| Meeting |
TMS Specialty Congress 2025
|
| Symposium
|
3rd World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2025)
|
| Presentation Title |
Strongly Physics-Constrained Neural Networks for Mechanical Superresolution |
| Author(s) |
Vivek Oommen, Andreas Euan Robertson, George Em Karniadakis, Remi Dingreville |
| On-Site Speaker (Planned) |
Vivek Oommen |
| Abstract Scope |
Neural operators facilitate the efficient estimation of the solution fields to partial differential equations. However, these methods are restricted by the availability of sufficient data for training. This problem is especially pronounced in crystal mechanics applications because at high resolutions the computational cost of running simulations becomes overwhelming. We propose a framework for multi-resolution training of neural operators in crystal mechanics, where the training is only supervised with low-resolution datasets generated from acceptably cheap simulations. To fill in the missing information, we introduce a new variant of neural operators: Structure- Preserving UNets. These networks are strongly physics-constrained: the deformation compatibility and stress equilibrium PDEs are directly incorporated into the architecture. In this talk, we present the novel architecture that makes physically consistent and computationally efficient predictions at high resolution. To understand the proposed framework’s strengths and weaknesses, we compare it against traditional softly constrained physics-informed neural networks. |
| Proceedings Inclusion? |
Undecided |