Abstract: The wetting deformation and the associated potential cracking risks pose a significant threat to the structural integrity of core wall rockfill dam. Current inversion methods for wetting parameters are predominantly deterministic, which neglect the effects of factors such as excavation disturbances, construction variability, and operational environment, leading to biased inversion results. In this study, taking the Pubugou dam as an example, a Bayesian framework-based probabilistic inversion method for wetting parameters is proposed. First, a statistical model is employed to decouple the settlement observations and extract the wetting-induced deformation component. Then, a Jaya-XGBoost model is developed as an efficient surrogate to replace the numerical model and perform sensitivity analysis. Finally, the surrogate model is coupled with DREAM(zs) sampling algorithm to achieve probabilistic inference and uncertainty quantification of the wetting parameters. The results demonstrate that the proposed method effectively overcomes the issues of non-uniqueness and local optima in deterministic inversion approaches and reduce parameter uncertainty. Moreover, the wetting deformation computed based on the posterior mean shows a smaller deviation from the observed monitoring data. This study provides a more accurate and reliable approach for parameter inversion of earth-rockfill dams.