Microscale strain field predictions from grain microstructure of polycrystalline metals using fully convolutional networks

Under deformation, the heterogeneous microstructure of polycrystalline metals generates complex strain variations at the microscale, which ultimately control failure mechanisms. Here, we train a fully convolutional network (FCN) on numerical datasets generated by crystal plasticity finite element simulations (CPFEMs) to predict the two-dimensional (2D) patterns of strain field variations (output) from grain orientation information (input) at the microscale, across a large subset of grain morphologies. Previously applied FCN architectures have correctly predicted the general patterns of strain distributions, but with performance that saturates quickly with increasing size of the training dataset. We overcome this limitation by augmenting the traditional convolution architecture with modern architectural elements such as skip connections, depth-wise separable convolutions, residual functions, and inverted bottleneck convolution modules, reducing the number of trainable parameters and floating-point operations by 88% and 77%, respectively. Our FCN architecture, trained on predominantly equiaxed grains with a fixed (lognormal) distribution of grain sizes under a small subset of macroscopic strain states, is capable of interpolation and limited extrapolation to other strain states. Its ability to predict the microscale strain patterns across a wide range of grain sizes, grain distributions, and grain shapes without retraining, further suggests its generalizability to different grain architectures. Finally, we discuss the utility of transfer learning to reduce the amount of training data required to adapt the FCN to materials with different stress–strain response.