Artificial intelligence resolves transboundary water conflicts under climate uncertainty

Sustainable transboundary water management is increasingly compromised by climate-induced deep uncertainty, as traditional open-loop strategies fail to adapt to non-stationary hydrological shifts and physical propagation time-lags. To overcome these structural rigidities, this study proposes a Physics-Informed Multi-Agent Reinforcement Learning (PI-MARL) framework. Operating as a closed-loop controller, it maps real-time basin states-explicitly accounting for physical routing delays-to optimal release decisions. Empirical validation in the Yarlung Tsangpo-Brahmaputra (YTB) River Basin reveals that AI agents autonomously learn cooperative, pre-emptive release strategies to achieve spatio-temporal risk substitution. Crucially, this robust coordination emerges through decentralized real-time execution, provided riparians agree upon a hybrid reward structure that internalizes transboundary risks. This effectively shifts the hydro-political challenge from operational micro-management to objective macro-negotiation. Rigorous stress testing under calibrated stochastic extreme events confirms that this AI-driven approach significantly outperforms traditional static baselines, reducing downstream flood peaks by 16.3% and elevating system reliability to 99.2%. By quantifying the value of information flow underpinning this mechanism, our findings provide a scientific foundation for transitioning transboundary governance from rigid quotas to adaptive, algorithmic real-time control.