Design of an interpretable Convolutional Neural Network for stress concentration prediction in rough surfaces

Highlights: • A CNN was trained to predict stress concentrations in rough surfaces. • A "viewport" was integrated in the CNN architecture for interpretability. • The CNN performed better than simple linear correlation analysis. • A CNN is a valuable tool for reducing computational cost of mechanical models. -- Abstract: We present the application of a Convolutional Neural Network (CNN) to relate stress concentrations to surface roughness. Stress concentrations at the low points of rough surfaces are one of the primary causes of fatigue crack initiation but there is no generally accepted method for analyzing rough surfaces to predict crack initiation. Synthetically generated rough surfaces, instantiated in a mechanical model allow for the simulation of stress concentrations, creating a database of surface images and corresponding mechanical data. In this work, the CNN is designed and trained to interpret a height map of a surface and, from that data, to predict the stress concentrations created by the surface. Using a simple architecture, the CNN achieved R² = 0.75 in prediction for test images, i.e., those not used in training. This CNN can be adapted for experimental surfaces thus creating a new and straightforward tool for prediction of crack initiation. Considerable care was taken to minimize the complexity of the CNN architecture and to make it interpretable via viewports.