Predictive Modeling of Real Contact Area on Rough Surfaces Using Deep Artificial Neural Network

Abstract

A computational approach employing the boundary element method (BEM) and a deep artificial neural network (ANN) is proposed to accurately predict the real contact area (RCA) for rough surfaces. A BEM-generated dataset is utilized to develop an optimized ANN model for the prediction of RCA based on surface topography parameters and applied load. The Bayesian optimized ANN model, employing a logsig transfer function in hidden layers and having a 5-44-44-1 architecture, predicts the RCA with a mean error of ≈6%. Shapley values are utilized for global sensitivity analysis, revealing that applied load and surface roughness are the most influential factors affecting RCA. The generalization capability of the ANN model is validated through comparisons with predictions on test data, demonstrating superior accuracy and computational efficiency over traditional numerical methods. The influence of surface topography parameters reveals that the higher surface roughness reduces RCA due to increased asperity interaction, while higher skewness promotes a larger RCA by increasing the number of contact points. Kurtosis influences RCA nonlinearly, with higher kurtosis yielding lower RCA at low loads but higher RCA at high loads due to asperity distribution characteristics. The optimized deep neural network model not only predicts the real contact area with high accuracy but also is significantly faster than boundary element simulations, conclusively demonstrating its potential to accelerate the design and optimization process in rough surface contact mechanics.