Prediction of elastic stresses in porous materials using fully convolutional networks

Abstract Machine learning (ML) models enable exploration of vast structural space faster than the traditional methods, such as finite element method (FEM). This makes ML models suitable for stochastic fracture problems in brittle porous materials. In this work, fully convolutional networks (FCNs) were trained to predict stress and stress concentration factor distributions in two-dimensional isotropic elastic materials with uniform porosity. We show that even with downsampled data, FCN models predict the stress distributions for a given porous structure. FCN predicted stress concentration factors 10,000 times faster than the FEM simulations. The FCN-predicted stresses combined with fracture mechanics captured the effect of porosity on the strength of porous glass. Increasing variations in pore size increased the variations in fracture strength. Furthermore, the FCN model predicts the pore configurations with the lowest and highest stresses from a set of structures, enabling ML optimization of porous microstructures for increased reliability.