Neural networks for groundwater prediction: a global literature review

Abstract

Groundwater is essential for water and food security, yet its prediction remains challenging because aquifer systems are nonlinear and monitoring data are often incomplete. This study aimed to determine the global state of the art in neural network applications for groundwater prediction through a systematic literature review. Publications from 2009 to May 2025 were retrieved from Scopus and Web of Science, yielding 359 records, of which 150 met the inclusion criteria. Data were organized and analyzed using Microsoft Excel, while VOSviewer 1.6.19 was used to characterize publication trends and keyword co-occurrence. India and Iran showed the highest scientific output. Long short-term memory (LSTM) models and hybrid architectures generally reported the strongest predictive performance, with several studies achieving root mean square error (RMSE) below 0.1 m and coefficients of determination (R²) above 0.95. Wavelet, empirical mode decomposition (EMD), and variational mode decomposition (VMD), together with satellite observations, improved performance in data-limited settings. The review indicates that future research should prioritize interpretable, transferable, and computationally efficient models that integrate neural networks with hydrogeological principles to support sustainable groundwater management.