GraphRAG Applied to Earthquake Catalogs for Knowledge Graph Construction
A recent study introduces the inaugural systematic use of graph-based retrieval augmented generation (GraphRAG) applied to unprocessed, tabular earthquake catalog data. This innovative pipeline constructs fully structured, queryable knowledge graphs for three distinct catalogs: the reservoir-adjacent swarm, the 2019 Ridgecrest tectonic sequence, and the Mw7.4 aftershock sequence in Maduo from 2021. By automating data structuring, this method facilitates responses to open-ended inquiries regarding earthquake sequences, moving beyond the constraints of fixed spatiotemporal frameworks and subjective expert analysis. The researchers conducted thorough evaluations against ground truth derived from catalogs and a rule-based reference graph, identifying failure modes with four prompts informed by seismology, addressing the surge in event counts in contemporary catalogs due to deep learning detectors and phase pickers.
Key facts
- First systematic application of GraphRAG to raw, tabular earthquake catalog records
- Three catalogs analyzed: reservoir-adjacent swarm, 2019 Ridgecrest sequence, 2021 Maduo Mw7.4 aftershock sequence
- Pipeline builds structurally complete, queryable knowledge graphs without manual data structuring
- Evaluation against catalog-derived ground truth and rule-based reference graph
- Four seismology-informed prompts used in evaluation
- Addresses increased event numbers from deep learning detectors and phase pickers
- Enables answering open-ended questions about earthquake sequences
Entities
Locations
- Ridgecrest
- Maduo