Graph Neural Networks for Contact Between Deformable Bodies

DOI: 10.26153/tsw/55894

Authors: Vijay Dubey, Collin E. Haese, Jan Fuhg, Manuel K. Rausch, Osman Gültekin

Abstract: Traditional methods, such as finite elements, are compute- and time-intensive when used to solve for displacements in nonlinear continuum mechanics, especially when multi-body contacts are involved. This is a roadblock for applications where such simulations need to be completed quickly. Machine learning can alleviate this. We present an implementation of Graph Neural Network for accelerating such simulations.

We utilized training data based on past simulations done using the finite-element method. This can be done due to the natural representation of meshes as graphs. The architecture employs multiple rounds of message passing based on encoded nodal and edge features (Dalton et al., 2022). Two types of edges are defined based on either connectivity in mesh or proximity in physical space (Pfaff et al., 2022). The learning is done through a composite loss term consisting of data-based dynamic losses arising from predictions of nodal accelerations and a self-supervised contact loss term. The contact loss computation is made tractable by using filters like bounding-boxes, sweep direction, vertex-face, and edge-edge tests to cull collision pairs (Zhu et al.,2022).

The framework differs from those in the past as it can be applied to contact between deformable bodies, unlike tool-object interactions, where the tool is considered rigid. We present the application of this framework to an example problem that demonstrates its capability to handle contact. Results on error statistics are also presented.

Keywords: Computational mechanics

How to Cite

@inproceedings{dubey_graph_2024,
	address = {Austin TX USA},
	title = {Graph {Neural} {Networks} for {Contact} {Between} {Deformable} {Bodies}},
	url = {https://hdl.handle.net/2152/129390},
	doi = {https://doi.org/10.26153/tsw/55894},
	booktitle = {TACCSTER Proceedings},
	author = {Dubey, Vijay and Haese, Collin E. and Fuhg, Jan and Rausch, Manuel K. and Gültekin, Osman},
	year = {2024},
}