Explainable Deep Learning for Settlement Prediction and Risk Classification of Transportation Infrastructure on Soil Foundations

This study developed an integrated framework for predicting settlement and classifying serviceability risk of transportation infrastructure on soil foundations. The dataset combined 418 finite element method (FEM) simulations covering 19 parametric soil scenarios under 22 incremental load steps, generated with the extended Mohr–Coulomb constitutive model. Eight algorithms — linear regression, tree-based ensembles, and deep learning architectures for tabular data — were benchmarked under a Group K-Fold protocol grouped by soil scenario and tested on an out-of-distribution set holding out two scenarios at the extremes of the effective elastic modulus. A three-class risk classifier was derived from Eurocode 7 and AASHTO LRFD serviceability thresholds. SHapley Additive exPlanations (SHAP) values from a tree-based and a deep learning model were cross-compared. The best surrogate achieved a root-mean-square error of 19.4 mm and a coefficient of determination above 0.99 on the log scale within distribution. The classifier reached a macro F1 of 0.98 in cross-validation and 0.88 on the out-of-distribution set, showing that risk zoning was more robust to extrapolation than absolute settlement estimates. SHAP analysis identified applied load as the dominant predictor, while soil-parameter rankings diverged across model families because of severe multicollinearity.