Investigation of As-Constructed Air Void Influence on Fatigue Cracking of Asphalt Pavement Using Explainable AI

Abstract

As-constructed air void is a critical control parameter in road construction, which affects the fatigue resistance of asphalt pavement. This study extracted and rigorously cleaned the data from the long-term pavement performance database to develop an artificial neural network (ANN) model for predicting fatigue cracking (FC). In addition, a series of benchmark experiments were conducted to investigate the effects of principal component analysis (PCA) cumulative variance thresholds and ANN architectures on model prediction accuracy and stability. An ANN architecture of 10-25-5-1 combined with a PCA cumulative variance of 90% showed the highest prediction accuracy for the long-term evolution of fatigue performance of asphalt pavements ( R 2 = 0.95 ). The SHAP-based global explanation found that the permutation of as-constructed air void resulted in significant changes in the loss function of the FC model, confirming that the as-constructed air void was a highly influential predictor for FC of asphalt pavement. Based on the developed ANN model, the parameter sensitivity analysis concluded that reducing the as-constructed air void content from 9% to 3% enhanced the fatigue resistance of asphalt pavement by 46.8%. The high precision ANN model developed in this study can provide quantitative support for maintenance planning and resource optimization.