Experimental Characterization of Various Mechanical Properties of the RAP Mixed Granular Blends and Modelling of Their Strain Accumulation Behavior Using Soft Computing Techniques of MLR, ANN, and RF

Abstract The use of reclaimed asphalt pavement (RAP) in various layers of flexible pavement offers noteworthy environmental and economic benefits, but its long-term performance under diverse environmental conditions remains insufficiently documented. This study evaluated the resilient modulus (M R ), accumulative strain, and permeability of RAP–aggregate blends under different stress states, temperatures, freeze–thaw (F–T) cycles, moisture levels, and loading characteristics. Additionally, it examined predictive modeling algorithms for accumulative strain. Experimental results showed that M R increased with RAP content, with more pronounced gains at bulk stresses > 250 kPa. Under drained testing conditions, F–T cycles improved M R by 10–20%, while higher moisture contents and elevated temperatures reduced stiffness. Permeability was reduced by about 50% at 20% RAP (0.00121 m/s to 0.0006 m/s). Longer loading wavelengths reduced M R by 46%, and triangular pulses generated the stiffest response. Accumulative strains increased noticeably with RAP content, reaching 1.67, 2.9, 5.0, and 8.0 times higher than virgin aggregates at 20%, 40%, 60%, and 80% RAP, respectively. Machine learning models estimated accumulative strain with varying success. Multiple linear regression (MLR) provided stable but reduced accuracy (R² ≈ 0.82), while random forests (RF) achieved high training accuracy but showed poor generalization. Artificial neural networks (ANN) delivered the best performance (R² up to 0.976, MSE ≤ 0.006), demonstrating robustness across validation schemes. The results emphasize that incorporating 20–30% RAP achieves a practical balance between stiffness and deformation, while higher RAP levels may require the use of stabilization agents for long-term durability.