Did They Feel It? Improving our Understanding of Earthquakes and Earthquake Effects using Legacy Macroseismic Data

Susan E. Hough, Aarnav Agrawal, Lori Dengler, William L. Ellsworth, Lijam Hajos, Margaret Hellweg, Khant Nyi Hlaing, S. Mostafa Mousavi, Robert McPherson, & Clara E. Yoon

Submitted September 7, 2025, SCEC Contribution #14473, 2025 SCEC Annual Meeting Poster #TBD

Even in a big-data era, we have limited instrumental seismic data for large earthquakes at close distances, which are critical for improving resilience of the built environment. Instrumental data can be augmented with empirical ground motion relations, physics-based simulations, and macroseismic data for notable past earthquakes for which few or no instrumental data exist. Starting in the 1920s, macroseismic surveys for earthquakes in the United States were collected by a succession of government agencies, using a questionnaire similar to that used by the current US Geological Survey Did You Feel It? system. Detailed responses were summarized in a series of US Coast and Geodetic Survey (USCGS) Reports now available online as flat PDFs. The MW 6.5 21 December 1954 Fickle Hill, California earthquake illustrates the potential value of what we dub Did They Feel it? (DTFI) data, providing data for a well-constrained ShakeMap representation of the event. DTFI data support results from analysis of early instrumental data (Hellweg et al., 2025), including the depth and suggestion of a high-stress-drop source. Because manual data mining and interpretation of DTFI and other macroseismic data is tedious, companion posters (Hlaing et al., 2025; Agrawal et al., 2025) present innovative approaches using Large Language Models (LLMs, or “AI”) to mine macroseismic data and assess intensities. The ML 5.3 22 March 1957 Daly City, California earthquake, which occurred on or very near where the San Andreas fault goes offshore from the San Francisco peninsula, provides a proof-of-concept (Agrawal et al., 2025). For this event DTFI accounts are available from over 2400 locations, 1973 with reported effects. Although the problem is tailor-made for LLMs, Agrawal et al. (2025) discuss a number of issues that arise. AI-generated MMI assignments, MMI-AI, are generally consistent with assignments in the USCGS report. Omitting locations where shaking was reportedly not felt (MMI 1), average AI assignments are ~0.26 units lower than USCGS MMIs. Much of the discrepancy can be attributed to the USCGS practice of assigning whole unit MMI values. Our detailed intensity distribution for the Daly City earthquake illustrates the potential for DTFI data to improve our understanding of notable 20th century US earthquakes (also see Agrawal et al., 2025).

Key Words
Macroseismology, Large Language Models

Citation
Hough, S. E., Agrawal, A., Dengler, L., Ellsworth, W. L., Hajos, L., Hellweg, M., Hlaing, K., Mousavi, S., McPherson, R., & Yoon, C. E. (2025, 09). Did They Feel It? Improving our Understanding of Earthquakes and Earthquake Effects using Legacy Macroseismic Data. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Ground Motions (GM)