What question did this study set out to answer?

March 2, 2026Open Access

Hybrid ML–XAI Framework for Predicting and Interpreting the Strength of Lime–Silica Fume Stabilized Clay for Sustainable Construction Applications

Puntos clave

To develop a predictive framework for the strength of lime-silica fume stabilized clay to support sustainable construction.
Conducted 210 UCS tests across 14 mix designs and three curing periods.
Measured index and compaction properties of kaolin clay soil stabilized with QL and SF.
Developed multiple ML models to optimize mix design using input variables like Atterberg limits and compaction parameters.
Applied PSO to tune hyperparameters of the models.
The QL–SF mixture achieved the highest UCS of 2783.8 kPa at 28 days, a 6.8-fold increase over untreated clay.
CatBoost-PSO and back-propagation neural networks exhibited R2 ≈ 0.97 and RMSE ≈ 105 kPa in testing.
SHAP analysis revealed QL content, plasticity index, and curing time as key predictors for UCS.

Resumen

This study presents an advanced experimental–computational framework for the characterization and performance evaluation of low-plasticity kaolin clay soil (CL) stabilized with quicklime (QL) and silica fume (SF), aiming to support sustainable construction and ground improvement applications. A comprehensive laboratory program was conducted, comprising 210 unconfined compressive strength (UCS) tests across 14 mix designs and three curing periods (3, 7, and 28 days), alongside index and compaction property measurements. The results show that stabilization decreases plasticity index (PI) and maximum dry density. The QL–SF system showed a synergistic effect, with QL3–SF7 mixture achieving the highest UCS (2783.8 kPa at 28 days), a 6.8-fold increase over untreated clay within the tested range. To enable predictive evaluation and mix optimization, multiple machine learning (ML) models were developed using eight input variables, including Atterberg limits and compaction parameters for each stabilized mixture, along with stabilizer contents and curing time, with hyperparameters tuned via particle swarm optimization (PSO). Among the evaluated models, CatBoost-PSO and back-propagation neural networks delivered the highest generalization performance on the independent testing dataset (R2 ≈ 0.97; RMSE ≈ 105 kPa over a UCS range of 408.88–2783.8 kPa). To enhance interpretability and engineering reliability, explainable artificial intelligence (XAI) using SHAP was employed to quantify feature influence and verify physical consistency. SHAP analysis identified QL content, PI, and curing duration as dominant predictors, and showed that SF contribution depends on its balance with available calcium from QL. Overall, the proposed ML–XAI framework provides a transparent decision-support approach for performance-driven design of chemically stabilized clay materials while reducing reliance on extensive trial-and-error laboratory testing.

Leer artículo completoexternamente

Preguntar a la IA

Me gusta

Guardar

Ver artículo completo