Ensemble machine learning method for δ¹⁸O prediction in groundwater

Groundwater δ¹⁸O interpretation in arid regions is challenging due to complex recharge, salinization, and limited datasets. We applied ensemble machine learning, Random Forest, Lasso, and Gradient Boosting, to predict δ¹⁸O using hydrochemical and spatial variables in the mid-Nile Valley. Under random splitting, RFR performed best (R² = 0.83), followed by Lasso (0.81) and GBR (0.80). Longitude, chloride, potassium, and magnesium emerged as the main predictors, consistent with hydrochemical evidence. But performance dropped under spatial cross-validation, a more realistic test: RFR R² = 0.50, Lasso = 0.45, GBR = 0.41. So, part of the predictive skill comes from spatial structure, not transferable relationships. A sensitivity test without longitude showed that hydrochemical variables alone explain about 40% of the δ¹⁸O variance (R² = 0.405). Within the data-rich range (-3‰ to -1‰, n = 54), the model achieved a mean absolute residual of 0.70‰ and a 95% prediction interval of ± 2.26‰. Outside this range, for Eocene end-members (n = 7) and Quaternary enriched samples (n ≈ 6), uncertainty increased substantially. Machine learning is useful for isotope prediction in data-scarce regions, but results must be interpreted with caution. Spatial effects matter. The model is best suited for site-specific analysis within its reliable range. Future work should focus on collecting more end-member samples to bridge the isotopic gap.