Boosting models for predicting soil compression index of Gulf Coast soils with an interpretability framework

Accurate estimation of the compression index (cc) is essential in geotechnical engineering, as it directly influences settlement predictions and design decisions for structures on soft soils. This study explores the use of machine learning to estimate cc from routinely available geotechnical parameters of fine-grained soils. A large dataset of 2135 samples from clayey soils in the US Gulf Coast region was used to compare the performance of five boosting algorithms: AdaBoost, Gradient Boosting, XGBoost, LightGBM, and CatBoost. CatBoost and Gradient Boosting achieved the highest mean coefficients of determination (R2 = 0.923) and the lowest root mean square errors (RMSE ≈ 0.22). These two models were subsequently selected for hyperparameter tuning via a Genetic Algorithm. After optimization, Gradient Boosting yielded the best test set performance (R2 = 0.948, RMSE = 0.157). To enhance interpretability, the Descriptive mAchine Learning EXplanations (DALEX) framework was applied, providing feature importance via dropout loss, Accumulated Local Effects, and Break Down decomposition of individual predictions into variable contributions and interactions. The results revealed that initial void ratio, natural water content, and dry unit weight were the most influential predictors. These findings demonstrate that boosting models combined with interpretability techniques enable reliable and transparent estimation of geotechnical parameters.