Experimental evaluation and machine learning-based prediction of bituminous concrete performance using the Marshall method

This study evaluated Bituminous Concrete (BC) mixes using Viscosity-Graded (VG) bitumen (VG10, VG30, VG40) through the Marshall method, analyzing stability, flow, voids, unit weight, and Voids Filled with Bitumen (VFB). A total of 48 specimens were tested, and machine learning models, Random Forest (RF) and Generalized Regression Neural Network (GRNN), were developed to predict performance based on bitumen content and grade. Performance metrics (R 2 (coefficient of determination), RMSE (root mean square error), MAE (mean absolute error)) were reported as the mean across 5 independent cross-validation folds. It was observed that GRNN achieved high accuracy (R 2 : 0.952 for stability; 0.979 for VFB). RF showed moderate but consistent performance (R 2 : 0.561 to 0.725). While GRNN demonstrated a strong interpolation capability within this experimental range, however its performance should be validated on independent datasets before generalization. The results suggested that VG30 offered a balanced performance, but optimal selection depends on traffic and environmental conditions. Although, these models cannot replace standardized laboratory testing, but these can provide a data-informed tool for preliminary mix assessment, that may reduce reliance on repetitive trials in pavement engineering applications.