Prediction of ground surface settlement induced by foundation pit dewatering using FOA-optimized gaussian process regression

To address the limitations of Gaussian process regression (GPR) in predicting ground surface settlement induced by foundation pit dewatering, namely inadequate predictive accuracy, limited generalizability, and reliance on empirical kernel selection, this study proposes a novel predictive framework, FOA-GPR. The model employs the fruit fly optimization algorithm (FOA) to optimize the weight allocation of a composite kernel comprising radial basis function (RBF) and Matérn kernels, thereby capturing complex, nonlinear spatiotemporal dynamics. Using six multidimensional features, including water table drawdown, cutoff curtain depth, and permeability coefficient, the proposed FOA-GPR framework is systematically evaluated against conventional machine learning models (BPNN, SVM, and GPR) as well as an Empirical Spatial Attenuation Model (ESAM). The results show that FOA-GPR delivers outstanding predictive performance, reducing the mean absolute error (MAE) by 48.2% and 76.8% relative to BPNN and ESAM, respectively, and lowering the mean squared error (MSE) by 83.8% compared with BPNN. In addition, the model achieves a high coefficient of determination (R²) of 0.975 and improves peak probability density by 14.3% over standard GPR. Finally, Shapley Additive Explanations (SHAP) analysis indicates that water table drawdown and spatial distance to the sump well are the dominant mechanical drivers of settlement. Overall, the proposed framework demonstrates superior accuracy, robustness, and physical interpretability, offering a highly reliable tool for risk management in deep excavation projects.