Mapping microstructure to mechanical property by disentangling strengthening mechanism with deep learning

The physics encoded in materials microstructures are essential to predict mechanical properties. However, disentangling or representing such information using deep learning remains a long-standing challenge due to the complexity in both microstructures and surrogate models. Here, we present an approach that comprises image augmentation, self-supervised learning and regression to achieve interpretable representation and improved prediction model. We demonstrate the proposed strategy on a small dataset of diverse measured microstructures and yield strengths. The learned representation (latent variables) shows a Hall–Petch like relationship with yield strength, indicating the capture of fine grain strengthening mechanism. As a result, the model accuracy for target property is doubled when applying to test data. Our approach can be generalized to other scenarios to recognize key physics for correlating microstructures to properties where limited data is available.