Intelligent Asphalt Mixture Design: A Combined Supervised Machine Learning and Deep Reinforcement Learning Approach

Abstract

Traditional lab-based asphalt mixture design is a time-consuming and labor-intensive procedure. To accelerate traditional asphalt mixture design and reduce excessive reliance on human input, we developed an innovative and intelligent mix design framework based on a deep reinforcement learning (DRL) algorithm and machine learning (ML) predictive models. Specifically, ML predictive models of asphalt mixture volumetric and mechanical properties were established using data from 598 Marshall mix designs. Next, the action, state, reward, and environment—the basic components of a DRL problem—were refined in alignment with the Marshall mix design procedure. Subsequently, with the predictions of ML models, a typical DRL model—deep deterministic policy gradient (DDPG)—was trained to design an asphalt mixture to achieve maximum Marshall stability and minimum mixture cost. The sensitivity of the important parameters of the DDPG model to its performance was then analyzed. Finally, the performance of the trained DDPG agent was verified with eight design cases and a comparative evaluation. The results show that the DDPG agent successfully produced eight mixtures that not only meet the specifications but also have higher stability and lower costs than lab-designed mixtures using the same type of binder. The performance comparison between the DDPG model and current popular mix design optimization algorithms (genetic algorithm and particle swarm optimization) demonstrates that the DDPG model not only achieves comparable mix design optimization effectiveness but also higher computational efficiency than these metaheuristic algorithms.