Machine learning paradigms in natural and engineered water systems: From proof-of-concept to trustworthy deployment.

Machine learning is increasingly used to model and manage water systems, from rivers and aquifers to treatment plants and distribution networks. Yet many studies remain proof-of-concept: models are trained on sparse or siloed data, behave as black boxes and rarely connect to operational decisions. Here we review representative applications across natural and engineered water-system archetypes and propose a decision framework for choosing among mechanistic, data-driven, and hybrid models under data, physics and deployment constraints. We then highlight three directions for moving from prediction to trustworthy action: (1) physics-informed and explainable approaches that enforce conservation laws and clarify decision drivers; (2) integration with digital twins and reinforcement learning to enable safe, closed-loop decision support; and (3) graph neural networks and federated learning to represent networked processes and share information without centralizing sensitive data. Collectively, these advances can make machine learning a practical tool for resilient, sustainable water management.