Towards geothermally induced earthquake forecasting using California’s open data

Meggy Rossbach, Zhengfa Bi, Nori Nakata, & Emily E. Brodsky

Submitted August 30, 2026, SCEC Contribution #15618, 2026 SCEC Annual Meeting Poster #TBD

Energy production from geothermal fields is a growing field in the energy sector. Increase in production often comes along with increased induced seismicity, where operations are regulated by the retro-active traffic light system. While induced seismicity often scales with operational parameters such as injected volume there is no currently applied model to forecast seismicity. Recent research has utilized deep machine learning models as a mechanism to forecast seismicity in geothermal systems based on daily well-specific data at The Geysers and Utah FORGE. The proposed model shows success in forecasting the seismicity rate for the two locations. However, geothermal fields all over the globe show variations in characteristics and operational parameters. In order to incorporate this heterogeneity in characteristics we aim to find a more generalizable forecasting mechanism, starting with a linear model and ARMAX model to benchmark the initial deep learning model and expanding our future methods to, e.g. to meta-learning. We aim to keep computation moderate and application simple so it can be based on public or limited operational data available. We start this work with the public data from California’s geothermal fields. For initial testing and comparison to the deep learning model we start with The Geysers and Brawley. We intend to bridge the spatial and temporal scale from daily well-specific data to monthly field-specific data. Further, we plan to explore different strategies on how to transfer the learnings from one site to another on the incomplete monthly data.

Preliminary results show that ARMAX performs and transfers relatively well under simple error metrics, setting a baseline that deep learning should exceed to justify its computational cost. Beyond the accuracy of the forecast, we evaluate input feature importance and the temporal relationships between features to gain physical insight into the processes governing geothermally induced seismicity.

Citation
Rossbach, M., Bi, Z., Nakata, N., & Brodsky, E. E. (2026, 08). Towards geothermally induced earthquake forecasting using California’s open data. Poster Presentation at 2026 SCEC Annual Meeting.

Related Projects & Working Groups
Earthquake Forecasting and Predictability (EFP)