This paper investigates the application of Artificial Neural Networks (ANNs) for predicting the resilient modulus (Mr) of subgrade and subbase soils, which is a critical parameter in pavement design. Utilizing a dataset of 1683 Mr observations, the ANN model incorporates eight input variables, including soil gradation, plasticity, and stress conditions. The model was optimized using a quasi-Newton method, achieving high predictive accuracy, with a coefficient of determination (R2) of 0.9613 and low error rates for both selection and testing datasets. To further enhance model interpretability, SHAP (SHapley Additive exPlanations) analysis was conducted, revealing the significant influence of specific input parameters, such as saturation ratio, plasticity index and soil gradation, on Mr predictions. This study underscores the potential of ANNs as a practical tool for estimating resilient modulus, offering a reliable alternative to conventional laboratory testing methods. The findings suggest that integrating ANNs into pavement design processes can lead to more accurate predictions of pavement performance, ultimately supporting the development of more efficient and durable road structures.
Read full abstract