AI Agent Scaffolding Boosts Earth Science Simulation Success to 84%

A novel AI framework named Knowledge Infrastructure (KI) enhances the capacity of autonomous agents to conduct intricate Earth science simulations. In a benchmark involving 3,000 trials on coupled-hydrology models, agents utilizing KI produced physically plausible and verifiable outcomes in up to 84% of the trials, whereas those without KI achieved results in less than 40%. KI externalizes domain knowledge into validated modeling operators, staged protocols, and diagnostic recovery methods. Additionally, researchers created a Knowledge Dissection Toolkit (KDT) that autonomously generated KI for 117 more process-based models across 14 Earth-science fields. Across all 119 KIs, modeling choices and failure solutions aligned. This research aims to bridge the gap between sophisticated simulation models and communities vulnerable to climate risks and resource shortages, who currently lack the necessary expertise.