Mapping Rough Surface Heights to Tractions with Convolutional Neural Network

Abstract

ABSTRACT Convolution neural network (CNN) has the potential to become a valuable tool in solving mechanical engineering problems with high accuracy and short solving time requirements. CNN has been applied in topography-related studies in other disciplines and some mechanical engineering fields and has proven effective. This study focuses on validating CNN in acquiring responding surface traction profiles of non-conforming contact under normal load as an alternative to the finite element or boundary element method to achieve the same level of accuracy with less time consumption. Previous studies showed including physical information in designing neural networks provides more interpretability to the networks' prediction results and improves the network's prediction accuracy. Based on the findings, this CNN includes the Greenwood-Williamson contact model and recent findings in rough surface contact to focus on characterizing the traction response based on asperities in contact. The physical information is included as a detailed-defined section isolated from machine learning that processes the input to extract critical information based on contact theories before training. A preliminary CNN model with physics information included via input processing is designed and trained with multiple regression methods. After training with 400 samples and tuning, the network can predict surface responding traction under normal loads using the image-to-image regression method with around 90% accuracy. The prediction results are evaluated using eye-inspection and image correlation metrics.