Answers are grounded in this paper and cite their sources.
A dataset of 1111 cement-treated clay samples—varying in soil composition, mix proportions, and curing conditions—was used to predict unconfined compressive strength (UCS) using three machine learning (ML) models: gradient boosted regression tree (GBRT), support vector machine (SVM), and random forest (RF). GBRT outperformed the others, achieving the lowest mean absolute error ( MAE = 90.03 kPa) and root mean squared error ( RMSE = 171.4 kPa), followed by SVM ( MAE = 92.53 kPa, RMSE = 174.71 kPa). RF, while less accurate ( MAE = 132.71 kPa, RMSE = 231.55 kPa), avoided negative UCS predictions. Key predictors identified through feature importance analysis included curing time, cement-to-soil ratio, water-to-soil ratio, and curing temperature. Accumulated local effect (ALE) plots clarified variable influence and model behaviour. Feature selection results aligned with geotechnical domain knowledge, with minor discrepancies due to feature redundancy. Limitations include negative UCS predictions by GBRT and SVM at low strength ranges, and poor extrapolation by RF. Overall, the findings suggest that ensemble models, particularly GBRT, are effective tools for predicting UCS. These models can supplement and guide conventional design practices, enhancing reliability and decision-making in soil improvement projects.
No takes yet. Share an insight, caveat, or question.
A 2025 study studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: