Interpretable Prediction of Unconfined Compressive Strength of Geopolymer-Stabilized Clayey Soil Using Kolmogorov-Arnold Networks

Journal of Ocean Engineering and Marine Energy · Springer Nature · 2026-10-30 · DOI: 10.1007/s40098-026-01649-5

Аннотация

Predicting the 28-day unconfined compressive strength (UCS) of geopolymer-stabilized clayey soil from mix-design and soil-index variables is a challenging non-linear regression problem driven by concurrent geochemical mechanisms. Machine learning approaches, including gradient boosting, XGBoost, and artificial neural networks, have achieved high accuracy on this problem but still rely on post hoc tools such as SHAP or partial dependence plots for interpretation. This study evaluates Kolmogorov-Arnold Networks (KANs) an architecture that places learnable B-spline functions on network edges for UCS prediction using a compiled database of 297 specimens characterized by ten geotechnical and mix-design variables. A five-fold cross-validation (CV) study across 15 configurations, conducted without any use of the test set for model selection, identified KAN-XL [10, 12, 8, 1] with grid resolution G = 5 as optimal (CV R² = 0.946 ± 0.018; 3136 trainable parameters; 224 edge functions). On the independent test set ( n = 60), all performance metrics of KAN are comparable with the performance of the other models, with R² = 0.981, RMSE = 0.525 MPa, and MAE = 0.338 MPa, which is a performance metric numerically competitive with gradient boosting (R² = 0.979), XGBoost (0.978), ANN (0.968) and SVR (0.915). The KAN derived marginal response curves are physically consistent with C-A-S-H gel formation kinetics, optimum fly-ash dosage window and NaOH molarity activation threshold. The number of feature pairs ( n = 10) is small, however, and the lack of a statistically significant rank agreement between KAN gradient sensitivity and SHAP importance from the reference GB model (Spearman = 0.055, p = 0.881) confirms that the two metrics capture different aspects of input–output dependence.

Читать у издателя →

Похожие статьи