Prediction and uncertainty quantification of soil consolidation coefficient using multi-kernel Bayesian Gaussian process regression

The coefficient of consolidation (C_v) is a key parameter for time-dependent settlement assessment in soft-soil engineering, but conventional laboratory determination is costly and time-consuming for corridor-scale projects. This study develops a multi-kernel Bayesian Gaussian process regression (MBGPR) framework for accurate and uncertainty-aware C_v prediction using routinely measured soil properties. The workflow includes three stages: evidence-based feature selection using negative log evidence (NLE), optimization of 15 kernel combinations, and Bayesian hyperparameter inference via Markov chain Monte Carlo (MCMC) for predictive uncertainty quantification. Using the primary Lianhuai dataset (n = 54), the model identified a parsimonious four-parameter subset (specific gravity, plastic limit, liquidity index, and internal friction angle) and achieved strong performance (5-fold cross-validation: coefficient of determination R² = 0.909, root mean square error (RMSE) = 0.34 × 10⁻³ cm²/s; leave-one-out cross-validation: R² = 0.912, RMSE = 0.33 × 10⁻³ cm²/s). Validation on a merged same-region dataset (n = 131) and an external cross-region Guangxi dataset (n = 188) showed robust predictive capability within similar geological contexts, with near-nominal 95% interval coverage (96.2% and 95.2%, respectively). SHapley Additive exPlanations (SHAP) analysis was used to improve interpretability by quantifying feature contributions and interaction effects. The proposed framework provides an efficient and transparent tool for probabilistic C_v assessment in corridor-scale soft-soil investigations while reducing reliance on extensive consolidation testing.