Application of machine learning and optimization-based mesoscopic parameter inversion in numerical models of cement-stabilized macadam

Cement-stabilised macadam (CSM) is widely used in road engineering due to its high strength and durability. Constructing accurate numerical models is crucial for analysing failure mechanisms and improving the performance and service life of CSM. However, the accuracy of these models largely depends on the precision of mesoscopic parameter inversion. Existing inversion methods mainly rely on formula-based derivations and manual trial-and-error approaches, which are time-consuming and often limited in accuracy, typically below 70%. To address these limitations, this study employs machine learning to enhance both the efficiency and precision of inversion. A semi-circular bending model of CSM was established using the discrete element method (DEM) to generate a dataset of macroscopic and mesoscopic parameters. Six machine learning models were evaluated, with the backpropagation neural network (BPNN) yielding the best performance. BPNN was further integrated with a particle swarm optimisation (PSO) algorithm to perform mesoscopic parameter inversion. Validation against laboratory tests showed that the PSO-integrated BPNN method reduced simulation errors to less than 5%, significantly improving inversion accuracy. Additionally, SHAP analysis identified the key factors influencing macroscopic mechanical properties. This study proposes an efficient mesoscopic parameter inversion method, providing valuable insights for optimising CSM materials in engineering applications.