همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Soil Shear Strength Prediction Using Genetic-Optimized Boosting Machine Learning Models

نویسندگان :

Mohammadreza Ghadami¹ Ali Noorzad² Hamid Mohammadnezhad³

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی 3- دانشگاه شهید بهشتی

کلمات کلیدی :

Shear Soil Strength،Machine Learning،Genetic algorithm،SHAP Analysis

چکیده :

Shear strength represents a vital characteristic of soils, viewed as their inherent ability to withstand failure under the influence of forces applied to the soil structure. However, it is worthy to note that conventional laboratory-based methods for the determination of soil shear strength parameter are soil sampling, time-consuming, and limited in scalability. In this context, empirical methods such as Skempton (1957), Karlsson and Viberg (1967), and Wroth and Wood (1985) were employed, including coefficient of determination (R²), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and mean square error (MSE). Among these empirical approaches, the Skempton approach (1957) performed better than the other empirical methods, achieving superior results across the metrics 0.6684, 0.0496, 0.0353, 0.1074, and 0.0022. Therefore, in this study, it has been tried to utilize a novel hybrid data-driven approach developed with the integration of machine learning boosting algorithms and a genetic algorithm to enhance the prediction accuracy and interpretability in soil shear strength estimation. Five machine learning boosting models, including Boosted Regression Trees (BRT), XGBoost (Extreme Gradient Boosting), Light Gradient Boosting Machine (LightGBM), AdaBoost (Adaptive Boosting), and CatBoost (Categorical Boosting), have been utilized. The hyper-parameters of each model have been optimized using Genetic Algorithm (GA), and the models have been evaluated through five statistical metrics. The optimized CatBoost model outperform in comparison to other models and also the regression-based models (which have been used by other researchers) in testing phase (with a R² = 0.7976, RMSE = 0.0273, MAE = 0.0363, MAPE = 0.0837, and MSE = 0.0013). Also, in order to improve the estimation, a clear predictive equation has been developed by the hybrid CatBoost-GA model, which provides proper and practical formulation for engineering applications. Moreover, SHapley Additive exPlanations made it clear that the liquidity index is the important variable controlling shear strength, followed by the dry density index. Surprisingly, lower-importance variables, that is, the plasticity index and loam content, exhibited weak associations with small median SHAP analysis, while wet density, which has been emphasized in prior literature, was also found to be less effective. In the light of this subject, the proposed hybrid ML–GA framework presents a reliable, interpretable, and computationally efficient alternative to conventional laboratory testing for the soil shear strength prediction.