NGBoost-based probabilistic surrogate modeling for rockfill dam settlements considering rockfill spatial variability

Rockfill spatial variability identification based on conventional Monte-Carlo stochastic finite element analysis can be computationally intensive. An alternative approach is to construct an efficient surrogate model that can accurately estimate the probabilistic rockfill dam settlement (RDS) under specified rockfill spatial variability. In light of this, a probabilistic surrogate modeling method is proposed for efficient and effective mean prediction and uncertainty quantification of the probabilistic RDS. In this methodology, the key influencing rockfill material parameters are firstly identified considering the joint effect of rockfill subzones by global sensitivity analysis, which are subsequently characterized using a cross-correlated multi-parameter and multi-zone random field. Afterwards, Natural Gradient Boosting (NGBoost) algorithm is adopted for probabilistic surrogate modeling to establish the mapping relationship between rockfill material parameters and dam settlement probabilistic distribution. A novel prediction interval optimization (PIO)-driven hyperparameter tuning method is developed to enhance the NGBoost-based surrogate model, which takes an above-target prediction interval probability coverage (PICP) and a minimal prediction interval average relative width (PIARW) as optimization objectives. Then, model interpretation using SHapley Additive exPlanation (SHAP) method is conducted to further validate the model effectiveness. The implementation of the interpretable PIO-NGBoost model is demonstrated on a real-world rockfill dam. The results indicate that the PIO-NGBoost model can produce accurate mean prediction and effective uncertainty estimation simultaneously, and exhibits more satisfactory performance compared with other commonly used probabilistic models. Besides, the proposed model can still be effective under different rockfill spatial variability configurations with superior computational efficiency. This study provides an advanced means to achieve excellent performance in probabilistic surrogate modeling at a low computation cost.