<p>This paper addresses the prediction of spring discharge, a crucial water source for populations and industries, using advanced machine learning (ML) methods. Since underground water sources play a vital role in water supply, optimizing and predicting the best-performing resources is essential. The study employs the group method of data handling (GMDH) as the core model, optimized by three algorithms: mouth brooding fish (MBF), colliding bodies optimization (CBO), and teaching–learning-based optimization (TLBO), which simulates teacher-student interactions in education. The optimized ML-GMDHTLBO model achieved the highest accuracy with a Squared Pearson Correlation Coefficient (<i>R</i>²) of 0.999, indicating 99.9% reliability in predicting spring discharge during the testing phase. The results show consistent high performance across multiple springs (Lordegan, Deime, Deh Cheshme, and Dehghara), with minimal error variation between training and testing, demonstrating superior stability and strong generalization ability compared to other models such as ANN, ANFIS, and SVR. ML-GMDHTLBO also achieved the lowest RMSE and MAE values, confirming its high predictive accuracy. These findings suggest that careful input selection and parameter optimization significantly enhance performance in complex karst systems, highlighting the practical applicability of the model for water resource management in industrial and urban contexts.</p> Graphical abstract <p></p>