Process-informed feature engineering for explainable optimization-driven AI in groundwater-level prediction

• A process-informed BOXGBoost framework is developed for groundwater-level prediction. • Hydrologically meaningful lagged, rolling, and cumulative features enhance model performance. • Bayesian optimization improves model robustness through automated hyperparameter tuning. • SHAP-based analysis provides interpretable insights into dominant hydroclimatic drivers. • Temporal dependencies and hydrological memory are identified as key controls of groundwater dynamics.