Data-driven framework for tribological property interpolation in nano-silica/polyurethane clearcoats

Abstract Acrylic polyurethane clearcoats are widely used as protective coatings in automotive and industrial applications; however, their limited abrasion and erosion resistance reduces long-term durability under severe service conditions. In this study, a comparative machine learning (ML) framework was developed to interpolate the mechanical and tribological properties of nano-silica/acrylic polyurethane clearcoats as functions of nano-silica loading (0–6 wt%) and particle type (fumed vs. precipitated). Experimental data comprised ten samples representing five nano-silica loadings for two particle types. Three regression models—Gaussian process regression (GPR), random forest, and ridge regression—were systematically compared using grouped cross-validation (Leave-One-Group-Out) to prevent information leakage between paired formulations. Ridge regression achieved the best overall interpolation performance for pull-off strength (R² = 0.64), erosion resistance (R² = 0.83), and abrasion resistance (R² = 0.54), whereas GPR provided the highest accuracy for hardness prediction (R² = 0.92). None of the evaluated models produced reliable interpolation of the friction coefficient (R² < 0.2), and this property was therefore excluded from the predictive analysis. The experimental results indicated that approximately 4 wt% fumed nano-silica provided the most balanced overall combination of tribological properties within the investigated range, whereas 6 wt% precipitated nano-silica produced superior hardness and abrasion resistance. SEM observations qualitatively supported the observed wear mechanisms by revealing a transition from severe abrasive wear to smoother polishing-type wear at intermediate nano-silica loadings. The proposed framework is intended solely for interpolation within the investigated composition range and should not be extrapolated beyond the available experimental data.