High-precision hybrid metaheuristic–machine learning framework with GUI implementation for predicting Marshall Stability and indirect tensile stiffness modulus in asphalt mixtures

Abstract Accurate prediction of asphalt mixture performance is essential for improving pavement durability and optimizing mix design efficiency. This study proposes a unified hybrid machine learning framework for the simultaneous prediction of Marshall Stability and Indirect Tensile Stiffness Modulus (ITSM) at 20 °C and 30 °C, enabling comprehensive thermo-mechanical characterization under intermediate service temperatures. Five advanced hybrid models, MOGB-RC, CatBoost-GA, LightGBM-HHO, ANN-DE, and ELM-WOA, were systematically developed and comparatively evaluated using cumulative error analysis, multi-metric statistical assessment, relative improvement comparison, and Taylor diagram validation. The results demonstrate that boosting-based hybrid models consistently outperform neural-based approaches in terms of absolute error reduction, squared-error control, bias stability, and generalization capability across temperature conditions. The MOGB-RC model achieved the highest predictive accuracy for Marshall Stability, reducing testing MAE to approximately 0.12 kN and RMSE to 0.15 kN, representing up to 44 % lower error compared to ANN-DE and nearly 85 % lower maximum deviation than ELM-WOA. For ITSM at 20 °C, boosting-based hybrid models (MOGB-RC and CatBoost-GA) maintained stable performance across the mid-to-high stiffness range, while ANN-DE exhibited nearly double the maximum error (∼470 MPa) compared to LightGBM-HHO (∼240 MPa), confirming superior robustness of ensemble-based approaches. At 30 °C, inter-model performance gaps narrowed; however, MOGB-RC preserved the lowest RMSE (∼50 MPa) and limited worst-case deviation to nearly 21 % of the mean stiffness value, whereas ELM-WOA reached nearly 30 %, highlighting significant differences in thermal generalization capability. Additionally, a dedicated graphical user interface (GUI) was developed to operationalize the optimized hybrid model, enabling rapid thermo-mechanical performance prediction and transforming the framework into a practical decision-support tool for intelligent asphalt mix design.