A generalizable and interpretable deep learning model to improve the prediction accuracy of strain fields in grid composites

Abstract

Recently, the design of grid composites with superior mechanical properties has gained significant attention as a testbed for deep neural network (DNN)-based optimization methods. However, current designed DNN architectures are not specifically tailored for grid composites and thus show weak generalizability in exploring unseen configurations that stem away from the training datasets. Here, a multiscale kernel neural network (MNet) is proposed that can efficiently predict the strain field within a grid composite subject to an external loading. Predicting the strain field of a composite is especially important when it comes to understanding how the material will behave under loading. MNet enables accurate predictions of the strain field for completely new configurations in unseen domain, with a reduced mean absolute percentage error (MAPE) by 50% compared to a benchmark, U-Net as current state-of-the arts DNN architectures. In addition, results showed that MNet maintained superb performances with less than one-third of dataset, and can be applied to grid composites larger than the composite configurations used for the initial training. By investigating the inference mechanisms from the kernels of multiple sizes, our work revealed that the MNet can efficiently extract various spatial correlations from the material distribution.

Keywords

• Deep neural network
• Interpretable deep learning
• Composite design
• Strain field
• U-Net
• Material design