COMPUTING THE DYNAMIC MODULUS OF ASPHALT CONCRETE CONTAINING RAP BY USING AI MODEL: DEEP NEURAL NETWORKS WITH BAYESIAN OPTIMIZATION

Abstract

The dynamic modulus (|E*|) of Hot Mix Asphalt (HMA) is a fundamental mechanistic property that characterizes the viscoelastic behavior of asphalt mixtures under varying loading rates and temperatures and serves as a critical input in mechanistic empirical pavement design. Laboratory determination of |E*| is time-consuming, costly, and requires specialized equipment, limiting its practical applicability. Predictive models are widely used, but conventional regression-based approaches often lack accuracy and generalizability. To address these limitations, this study develops a machine learning (ML) based approach using Artificial Neural Network (ANN) to estimate the |E*| of HMA mixtures containing reclaimed asphalt pavement (RAP). The present study employed a dataset compiled from recently published research, comprising a total of 160 data points, which encompass cumulative percent retained on 3/8 inches sieve (P38), cumulative percent retained on No. 4 sieve (P4), percent passing No. 200 sieve (P200), air voids (Va), voids in mineral aggregate (VMA), voids filled with asphalt (VFA), effective asphalt content (V_beff), asphalt content (AC), dust to binder ratio (DP), testing temperature (T), loading frequency (F), binder replacement type RAP (BR_Type_RAP), and binder replacement percent (BR%) as predictor variables to develop a  Deep Neural Network (DNN) through BO for estimating the |E*| of HMA mixtures. The developed DNN model exhibited high prediction accuracy across all data subsets, with R² exceeding 0.99 across all data subsets, while MAE remained under 486 MPa and MAPE was below 9%. Sensitivity assessment identified that T and the P4 had the greatest influence on the predicted |E*| values.