Artificial intelligence models for predicting unconfined compressive strength of mixed soil types: focusing on clay and sand

Unconfined Compressive Strength (UCS) is a critical parameter in geotechnical engineering, influencing soil stability, foundation design, and load-bearing capacity. Traditional UCS prediction methods, such as Multiple Linear Regression (MLR), often struggle to capture the non-linear relationships inherent in mixed soil compositions. This study evaluates the effectiveness of Artificial Intelligence (AI)-based models, including Artificial Neural Networks (ANN), Support Vector Regression (SVR), and Random Forest Regression (RFR), in predicting UCS for clay-dominant and sand-dominant soils. A dataset of 100 soil samples from six geographically diverse regions across India was analyzed, incorporating key soil parameters such as clay content, sand content, liquid limit, plasticity index, and curing period. The models were assessed using R2, Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Prediction Interval (PI), and Index of Agreement (IOA). Among the AI models, RFR outperformed others with an R2 of 0.93 (training) and 0.84 (testing), a20 accuracy of 95%, and minimal error rates, demonstrating its superiority in UCS prediction. Sensitivity analysis identified clay content (28.7%) and curing period (22.1%) as the most influential factors affecting UCS, reinforcing their geotechnical significance. Regularization techniques such as dropout, batch normalization, and early stopping were implemented to prevent overfitting in ANN, ensuring model generalizability. To enhance interpretability, feature importance analysis and correlation analysis were conducted, allowing insights into how soil parameters influence UCS. The study also discusses potential advancements, including hybrid AI-geotechnical models, ensemble learning approaches, Bayesian optimization for hyperparameter tuning, and expansion of datasets with global soil data. The findings highlight the potential of AI-driven techniques as robust, scalable alternatives to traditional UCS prediction methods, with implications for soil classification, stability assessment, and foundation engineering.