Toward transparent groundwater contamination risk forecasting: Integrating causal discovery and Bayesian graph neural...

Groundwater contamination poses severe risks to ecosystems, agriculture, and public health. However, predictive modeling is limited by the complexity of environmental systems and the opacity of conventional machine learning models. This study introduces the Enhanced Bayesian Causal Graph Neural Network (EBC-GNN), a novel framework that integrates causal discovery via DYNOTEARS, spatiotemporal graph neural networks, and Bayesian uncertainty quantification to address these challenges. Leveraging heterogeneous, multi-source datasets, including EPA water quality records, land cover classifications, climate data, and industrial facility registries-the EBC-GNN identifies causal environmental drivers and produces robust, interpretable predictions with quantified uncertainty. Applied to California's Yolo and Tulare counties, the framework achieved a 70 % positive R² success rate, substantially outperforming conventional machine learning baselines. Beyond predictive improvements, the model uncovered policy-relevant insights highlighting the protective effects of wetlands and the elevated contamination risks associated with agricultural land use. By coupling interpretability with regulatory alignment, the EBC-GNN establishes a new benchmark for environmental risk modeling and decision support, offering a scalable and transparent tool for guiding sustainable groundwater management strategies.