Machine learning approach to the assessment and prediction of solid particle erosion of metals

Abstract

Solid particle erosion (SPE) is a tribological phenomenon in which a surface is impacted by a stream of particles, causing gradual removal of material. This poses significant challenges in aerospace, particularly when operating in harsh environments. Despite decades of data gathering and empirical model development, accurately predicting SPE remains challenging due to the complexity of the phenomenon and the variability in testing conditions. In this study, we compiled a database of over 1000 erosion tests on metals from existing studies and internal experiments, noting material properties, test conditions, and literature metadata. Machine learning (ML) models, including Random Forest, Neural Networks, Support Vector Regression, and XGBoost were employed to predict erosion rates. XGBoost was most performant, achieving a mean absolute error of 15–16 % on test data. Model performance was further validated by predicting results published in the ASTM G76 standard; predictions were within the interlaboratory standard deviation for tests at 70 m/s. Feature importance and partial dependence plots were used to evaluate the influence of different variables on erosion predictions. While particle velocity, particle size, and impact angle show the expected influence, features such as target density and Poisson’s ratio showed exaggerated effects due to their role in classifying outlier materials. These results show the promise of ML for SPE prediction across a range of conditions and suggest that the broader erosion literature is valuable for quantitative predictions, while also acknowledging limitations in the ML approach, particularly where data sparsity and feature correlations hinder the accurate assessment of feature influence.