Machine Learning-Driven Prediction and Interactive Nonlinear Analysis of Compaction Parameters for Fine-Grained Soils

Abstract

Compaction parameters of soil material, maximum dry density (MDD) and optimum moisture content (OMC), are critical control indicators for highway embankment construction. In this study, a dataset containing 199 compaction test results for fine-grained soils was collected. Using MDD and OMC as prediction targets, Random Forest (RF), Gradient Boosting (GB), Extreme Gradient Boosting (XGBoost), and Support Vector Regression (SVR) were developed and optimized using the Wild Horse Optimization (WHO) algorithm. Five universal evaluation metrics, combined with radar charts, were used to comprehensively compare the predictive performance of each model. Based on univariate sensitivity analysis and SHapley Additive exPlanations (SHAP) global feature interpretation, two-way partial dependence plots (PDPs) were applied to reveal the pairwise nonlinear relationships between soil physical indices and compaction indicators. The results demonstrate that WHO-tuned XGBoost achieves optimal comprehensive predictive performance for both MDD (test set R2 = 0.8226) and OMC (test set R2 = 0.7486), outperforming SVR, GB, and RF in terms of fitting accuracy and generalization under small-sample conditions. Plastic limit (PL) exerts a significant influence on compaction performance. By further comparing WHO, Particle Swarm Optimization (PSO), Gray Wolf Optimizer (GWO), Random Search, Bayesian Optimization and Grid Search on the optimal model, the applicability and feasibility of WHO-XGBoost in predicting compacted-soil compaction parameters were validated. The proposed data-driven compaction evaluation framework (WHO-XGBoost-SA-SHAP-PDPs) acts as an auxiliary tool to lower the workload and cost of laboratory Proctor tests, offering theoretical support and technical guidance for rapid refined embankment compaction control in green transportation infrastructure.