HESS Opinion: Beyond the Machine Learning Paradox in Hydrology: A Framework for Scientifically Useful Modeling

Abstract. Machine learning (ML) is rapidly transforming hydrologic modeling, delivering remarkable predictive capabilities across applications. Yet a fundamental question remains: does greater predictive performance necessarily advance hydrologic process understanding? While ML models can outperform process-based approaches, they often provide limited mechanistic insight, violate physical constraints, and struggle beyond their training data. We refer to this tension as the “Machine Learning Paradox in Hydrology”: the flexibility that enables ML to capture nonlinear catchment responses can also decouple predictive performance from mechanistic interpretability. This tension raises a broader challenge for hydrologic modeling: how can predictive models be used not only to forecast system behavior, but also to test, refine, and advance understanding of the processes that govern it? In this Perspective, we propose a framework for advancing data-driven hydrologic modeling based on five dimensions: predictive skill, physical plausibility, extrapolative robustness, process-diagnostic value, and theory-building capability. Using examples of ML underperformance across the conterminous United States (CONUS), we illustrate how model failures can reveal unresolved processes, missing information, and limitations in existing physical representations and theory. We then propose four complementary strategies to address the ML paradox: physics-informed ML, explainability, transfer learning, and residual structural learning. Such capabilities would transform hybrid models from predictive tools into process-informed reasoning systems that actively support hypothesis testing, model falsification, mechanistic interpretability, and theory development.