Research Summary: CDEP Agent: Connecting Meteorologically Detected Temporal Compound Events to Real-World Documentary Evidence

A new LLM-agent framework, CDEP Agent, has been developed to bridge the gap between climate science's identification of compound drought-to-extreme-precipitation (CDEP) events and their recognition in real-world early warning systems and post-event documentation. The framework, designed to be auditable, links meteorologically detected CDEP candidates to various real-world hazard and impact evidence sources, accommodating differences in spatial scales, temporal resolutions, and reporting conventions. Initial application in California identified numerous CDEP events, suggesting a method for improved event tracking and evidence integration.

Why it matters

This research is strategically important because it addresses a critical disconnect between scientific understanding of complex climate events and their practical recognition in hazard mitigation and response. By providing a tool to systematically link meteorological data with real-world impacts, it enables more robust risk assessment and could inform better preparedness strategies across various sectors impacted by extreme weather. This could lead to more effective resource allocation and resilience planning.

Key insights

• Compound drought-to-extreme-precipitation (CDEP) events are increasingly recognized in climate science as significant drivers of extreme impacts.
• There is an identified gap between meteorological detection of CDEP events and their documentation or advance warning in real-world contexts.
• The CDEP Agent is an auditable LLM-agent framework designed to connect meteorologically detected CDEP candidates with real-world documentary evidence.
• The framework addresses challenges posed by disparate data sources with varying spatial scales, temporal resolutions, and reporting conventions.
• A case study in California identified 408 candidate CDEP events from ERA5 observations during 2021-2025 using the CDEP Agent.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2608.28628