GMFAgent: domain knowledge-driven agent for ground motion field estimation

Rapid estimation of the post-earthquake ground motion field (GMF) is crucial for impact assessment and emergency management. However, existing research and software rely heavily on expert judgment for selecting and weighting ground motion prediction models (GMPMs). This dependence often leads to subjectivity and results in static weighting schemes that lack the flexibility for dynamic adjustment based on specific event characteristics. To address these limitations, this study introduces GMFAgent, an autonomous artificial intelligence agent system driven by large language models (LLMs) for GMF estimation. The agent is capable of adaptively generating spatial meshes around the epicenter based on real-time source parameters and fusing necessary elements from multi-source data. GMFAgent innovatively incorporates a retrieval-augmented generation (RAG) mechanism powered by a vector database. This allows the system to retrieve information from a specialized domain knowledge base, including technical manuscripts and peer-reviewed reports on GMPMs. By simulating the logic of expert review, the proposed agent autonomously determines the weights for different GMPMs and orchestrates the built-in OpenQuake Engine and custom models to execute calculations. A case study using a subduction interface scenario earthquake off the east coast of Japan demonstrates the system’s functionality and reliability. It is anticipated that GMFAgent will significantly enhance the level of automated decision-making in post-earthquake emergency response.