Predicting groundwater recharge potential across various physiographic divisions of Bangladesh using generative data augmentation

Study region The present study region exhibits six diverse physiographic divisions of Bangladesh. Study focus This study harnessed the potential of generative artificial intelligence (GenAI) to predict groundwater recharge across physiographic divisions, ranging from recent to old floodplains, terraces, and a depression area. A total of thirteen predictive variables, encompassing both surface and subsurface factors, and a recharge inventory from 227 piezometers estimated using the water table fluctuation method were used. Since classical machine-learning models often achieve limited accuracy when trained on small datasets, we deployed two GenAI models: Conditional Tabular Generative Adversarial Network (CTGAN) and Tabular Variational Autoencoder (TVAE) to augment the existing data and improve predictive performance. The generated synthetic observations reproduced the statistical structure of the original data and strengthened the downstream machine-learning predictions. New hydrological insights for the region This study provides clear insight into the applicability of machine-learning models for groundwater-recharge prediction under limited-data conditions. Following GenAI-based data augmentation (n = 10–150), relative test-R² gains ranged from 8.3% to 24.2% across the five regressors. Gradient Boosting with TVAE augmentation achieved the largest relative gain (+24.2%, n = 150), whereas XGBoost with CTGAN augmentation achieved the highest absolute test R² of 0.716 (n = 20). These model-level improvements were evaluated using the pooled dataset encompassing all physiographic divisions, with groundwater recharge expressed in millimetres (mm). Spatially, active floodplain areas exhibited the highest recharge potential, followed by terraces and older floodplains, whereas areas adjacent to the depression exhibited relatively low recharge potential. These results demonstrate the strategic value of GenAI for addressing data scarcity and advancing groundwater-recharge modelling.