Cross-model feature-importance analysis of soil properties for predicting optimum moisture content and maximum dry unit weight of fine-grained soils

This study evaluates the influence of routine soil index properties on the prediction of optimum moisture content (w_opt) and maximum dry unit weight (γ_dmax), which are the primary outcomes of the Proctor compaction test, using machine learning (ML) methods. A curated database of fine-grained soils (n = 465, drawn from 15 sources) included gravel content (GC), sand content (SC), fines content (FC), liquid limit (LL), plastic limit (PL), plasticity index (PI), specific gravity (Gs), w_opt, and γ_dmax. After correlation-based feature filtering, three models were developed: Generalized Additive Model (GAM), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). The training used nested cross-validation with Bayesian optimization, corresponding to an overall 80–20 train-test split. The model performance was evaluated using R², RMSE, MAE, MAPE, r, and the overfitting ratio calculated for the test set. For w_opt, the best GAM model achieved R² = 0.84 and RMSE = 2.16%, outperforming RF and XGBoost. For γ_dmax, the best GAM and XGBoost models reached R² = 0.79 and RMSE = 0.76 kN/m³, respectively. SHapley Additive exPlanations (SHAP), model-based importance scores, and single ablation analyses consistently identified LL and PL as the most influential predictors, and FC provided secondary contributions, while GC and Gs added little once LL and PL had been included. Moreover, paired-feature ablation confirmed the joint influence of LL and PL on the prediction. Overall, all three models predicted compaction parameters with good accuracy; however, GAM models achieved comparable or better predictive metric values than the ensembles (RF and XGBoost) while offering interpretability through plots linking soil indices with the predicted outcomes. This balance of accuracy and interpretability supports GAM as the preferred model for prediction modeling.