MPNN based graph networks as learnable physics engines for deformation and crack propagation in solid mechanics

Abstract

In this paper, the message passage neural networks (MPNN) based graph networks is proposed for learning and prediction of boundary loading problem in solid mechanics. As a learnable physics engine, the proposed model is designed to learn from historical data to understand the rules of the physical system behind the data, and then to make predictions. By learning the historical displacement field sampling data of 2D intact plate, the proposed model successfully realizes its continuous displacement field prediction. On this basis, the trained model also achieves a good prediction in a plate containing holes under the same loading condition but without training, which verifies the generalization ability of the proposed model. Finally, the crack propagation can be realized by learning the historical data of the plate with prefabricated cracks subjected to tension, which shows the excellent predict ability and exciting prospect of the proposed model.