Adaptive physics-informed neural network framework for PTFE wear prediction through Archard-guided physical embedding

Abstract

Accurate prediction of polytetrafluoroethylene (PTFE) wear volume from measured operating and surface-topography data is important for evaluating wear evolution in frictional components. However, traditional wear laws often use a simplified representation of the coupling between operating conditions and surface morphology, and purely data-driven models may have limited interpretability and generalization ability when only small datasets are available. In this study, we established a unified neural network framework to compare five physics-embedding strategies for predicting the wear volume of PTFE. The comparison included our proposed adaptive physics-informed model, PI-A, four representative variants, and a baseline artificial neural network (ANN) as a reference. A dataset comprising 168 samples was established through reciprocating wear tests using PTFE plates against AISI 440 C steel balls. The input feature set was determined by combining the Boruta-SHAP and recursive feature elimination methods. The results showed that PI-A achieved the highest predictive performance, with the mean test R2 increasing by 8.93% and the mean test MSE decreasing by 55.02% relative to the baseline ANN. Bootstrap-based instability analysis further indicated that PI-A had the narrowest instability range and the lowest MAPE instability. These results suggest that adaptive hidden-layer modulation of an Archard-derived descriptor is an effective method for incorporating physical prior information into PTFE wear prediction.