Wear rate prediction of PVD hard coatings using a hybrid Taguchi–RSM–machine learning framework with SHAP-based interpretability

Introduction. The wear behaviour of physical vapour deposition (PVD) hard coatings is difficult to predict because of the nonlinear relationship between coating chemistry, deposition thickness, operating temperature, and applied contact load. Multi-dimensional parameter mapping through exhaustive experimentation is both time-consuming and costly. The purpose of the work is to implement a low-data hybrid architecture that combines Taguchi design, response surface methodology (RSM), and machine learning (ML) to predict and optimize wear performance based on a small dataset (n = 28). Methods. Three coating types – AlTiN, CrN, and TiC – were tested at temperature range of 40–50°C, contact load range of 5–15 N, and coating thickness range of 2–4 µm. Analysis of variance (ANOVA) was performed to identify the most influential parameter affecting wear. A predictive wear model was developed using response surface methodology. Results and discussion. ANOVA revealed that load and coating chemistry are the most influential factors affecting wear. Temperature and thickness were not found to be significant within the studied range. A RSM model was found statistically significant for predicting contact load and coating chemistry with R2 = 0.901 (Adj-R2 = 0.834, RMSE = 0.0076). Random forest (RF) had highest generalisation performance (CV R2 = 0.755) and gradient boosting (GB) had the best foverall fit (R2 = 0.913) among the ML models tested using five-fold cross-validation. SHAP analysis indicated that coating chemistry formed the major contribution, then contact load, and little temperature contribution. Gradient boosting optimisation indicated that AlTiN at 4 μm thickness, 15 N load and 50 °C are the preferred settings and the expected wear rate is 0.010823 (mm3/Nm). The proposed framework demonstrates credible, interpretable wear forecasting using limitted experimental data.