Realizing adaptive governance of coupled climate-social systems: A deep reinforcement learning framework

The escalating pressure of global warming necessitates practical and effective climate governance policies to meet the urgent goals of the Paris Agreement . However, for conventional policy simulation models like integrated assessment models (IAMs), their reliance on predetermined scenarios makes it challenging to adequately address the substantial uncertainties inherent in climate-social system evolution. This study proposes an adaptive decision-making framework that integrates deep reinforcement learning (DRL) with a climate–social system model to identify governance strategies that can reduce the risk of transgressing planetary boundaries. In this framework, we operationalize Social Tipping Elements (STEs) as targeted actions. A reward function provides feedback to optimize these interventions, ensuring the system remains within critical planetary boundaries. Our results demonstrate that, compared to conventional static models, the framework discovers adaptive policies through dynamic learning, enabling real-time responses to evolving climate-social conditions. These adaptive policies exhibit an “early-strong intervention, mid-term moderation, and late-stage reinforcement” pattern that reduces planetary boundary overshoot time by 55 years while stabilizing global warming below 1.5°C by 2100. By varying governance objectives, we further reveal the trade-off mechanisms between competing climate and socioeconomic goals. Notably, the designed multi-objective reward function enables a synergistic balance across competing objectives, resulting in the shortest planetary boundary overshoot duration (15 years) among all evaluated scenarios. This framework overcomes the limitations of static simulations, offering a robust and interpretable tool to design adaptive strategies crucial for navigating the competing objectives of time-sensitive climate governance.