Can neural nets leverage modern surface rupture data to improve paleoseismic magnitude estimates?

Allison C. Schiffmaier, Yajaira De Haro, & Alba M. Rodriguez Padilla

Submitted September 7, 2025, SCEC Contribution #14785, 2025 SCEC Annual Meeting Poster #TBD

The magnitudes of paleo-earthquakes are useful for hazard assessment. Traditionally, these magnitudes are estimated based on individual displacements from trench data or geomorphic offsets, or from rupture length extents, based on empirical scaling relationships between these features and magnitude. As paleoseismic data improves and multiple data points become available for a single paleo-event, can magnitude estimation methods be improved to leverage the additional information available? We test whether machine learning models can improve predictions of paleo-earthquakes by training and validating on modern displacement and rupture extent data. Using the FDHI database, we trained models on raw/normalized displacement, rupture position, and rupture length with Python methodology on Jupyter Notebooks on VS Code. We evaluated performance across varying data densities and data types, tracking R², MAE, RMSE, and feature importance. We tested ML-based regression methods and neural nets, finding the latter outperforms regression methods. As expected, the magnitude predictions degrade with decreased data volume until 60% of the data has been removed, at which point the quality of the prediction steadies. These preliminary results indicate that different methods could provide the best magnitude estimates as a function of data volume and type. Future work will apply these methods on trench data and geomorphic offsets from California.

Key Words
Earthquake magnitude, Paleoseismology, Neural network, MLP, Machine Learning Models, Regression Models, Displacement, Rupture length, Model performance

Citation
Schiffmaier, A. C., De Haro, Y., & Rodriguez Padilla, A. M. (2025, 09). Can neural nets leverage modern surface rupture data to improve paleoseismic magnitude estimates?. Poster Presentation at 2025 SCEC Annual Meeting.

Related Projects & Working Groups
Earthquake Geology