From descriptive digital twins to closed-loop intelligent agents: Adaptive biogeochemical management in...

The increasing frequency of non-stationary hydrological extremes is outpacing the response capacity of conventional rule-based watershed management, leading to elevated risks of algal blooms and contaminant pulses in river-reservoir systems. Addressing these dynamic environmental risks necessitates a transition toward adaptive, closed-loop decision-making. Beyond ongoing challenges in data scarcity and model transferability, a critical operational bottleneck is the gap between reliable prediction and safe physical intervention. To address the translation of model outputs into physical control actions without violating safety boundaries, this review establishes a maturity framework linking artificial intelligence capabilities to the safe delegation of decision authority. Three core contributions are presented. First, physics-informed machine learning and spatiotemporal graph networks are synthesized to show how physical conservation laws overcome data scarcity and topological complexity. Second, the closed-loop decision layer is critically examined through targeted operational scenarios: virtual sandboxes for preemptive algal bloom mitigation, multi-agent reinforcement learning for dynamic flood coordination, and inverse modeling for rapid chemical spill containment. Third, deployment barriers are distilled into four foundational pillars: computational infrastructure, algorithmic robustness, cyber-physical security, and institutional accountability. Finally, an open benchmark ecosystem is proposed to operationalize this shift, advancing watershed digital twins from passive monitors into responsible intelligent agents.