Multi-Constitutive Neural Network for Large Deformation Poromechanics Problem

Proceedings of the Machine Learning and the Physical Sciences Workshop, 34th Conference on Neural Information Processing Systems (NeurIPS), Virtual (Online), 2020.

Q. Zhang, Y. Chen, and Z. Yang contributed equally to this work.

• Manuscript

• Poster

Summary

The multi-constitutive neural network (MCNN) addresses large-strain consolidation while allowing a single trained model to represent several constitutive laws. Instead of training a separate solver for each material law, the method adds a one-hot constitutive-law label to the dimensionless spatial and temporal coordinates $\hat{X}$ and $\hat{t}$ supplied to a deep neural network. The label selects the relevant governing equation during training and prediction. Demonstrations use three hyperelastic laws and compare the learned responses with finite-difference reference solutions and independently trained physics-informed networks. The shared model gives accurate consolidation and settlement predictions across the tested laws. In two of the three benchmark cases, it also achieves lower errors than the corresponding independent networks. These results suggest that jointly learning related governing equations can improve data efficiency and generalization while reducing the need for separate models.