Leakage-safe and uncertainty-aware surrogate modelling for mechanistic tensile strain prediction in flexible pavements

Mechanistic pavement analysis relies on accurate prediction of critical structural responses, particularly tensile strain at the bottom of the asphalt layer (εt), obtained from multilayer elastic analysis. Repeated mechanistic simulations, however, can become computationally demanding in large parametric studies, optimisation, and reliability assessment. This study develops and validates an engineering-oriented surrogate modelling framework for rapid prediction of KENPAVE-generated tensile strain within a predefined structural–loading domain. A canonical dataset of 2,076 simulations was evaluated using leakage-safe five-fold GroupKFold validation with 1,887 independent validation groups. Support Vector Regression with a radial basis function kernel provided the best performance, achieving an out-of-fold R² of 0.99833 and MAPE of 1.501%. The framework also integrates empirical residual-based uncertainty quantification and domain-of-applicability assessment. Prediction intervals achieved coverage close to the nominal 95% level in well-supported regions, while reduced reliability was identified in low-support regions. The surrogate substantially reduced computational time, enabling efficient parametric exploration, optimisation, and reliability analysis. The approach is intended to accelerate mechanistic response evaluation within the defined simulation domain rather than serve as a universal fatigue-performance prediction model.