A hybrid GPR framework with feature selection for enhanced prediction and explainability of subgrade soil strength

In order to overcome the drawbacks of the labour-intensive California Bearing Ratio (CBR) test for pavement design, a Hybrid Gaussian Process Regression (GPR) model for predicting subgrade soil strength was developed. The model is optimised using the Bayesian Tree Parzen Estimator and uses Recursive Feature Elimination for feature selection. Several machine learning techniques, such as Multi-Linear Regression (MLR), Multi-Layer Perceptron of Artificial Neural Networks (MLPNN), and Random Forest (RF), were used in the first stage of evaluation on a regional geotechnical database with over 1000 samples. In the second stage, performant hybrid models based on Dense Neural Network and tree-based ensembles models were used. All of the machine learning models utilised in this study were outperformed by the Hybrid GPR model, According to the trial results, the hybrid GPR model outperformed all the machine learning models used in this study. It achieved the highest accuracy with the mean squared error (MSE), mean absolute error (MAE), mean absolute percentage error (MAPE), and the determination coefficient (R²) presented in the following (training/testing) pairs: (0.0768/0.2974), (0.2107/0.4364), (5.32/13.79)%, and (0.9982/0.9935), respectively. The SHAP interpretability method is used to analyse the prediction outcomes in order to further investigate the "black box" problem of the prediction model. According to the SHAP study, features related to dry density, gravel content, and optimum moisture content appear to be the main factors influencing subgrade soil strength based on CBR values. The results of the study show that the Hybrid Gaussian Process Regression model effectively captures the intricate nonlinear interactions between the feature parameters and offers a crucial foundation for CBR value estimation.