Sadegh Karimpouli published a paper in Geophysical Journal International entitled “Forecasting induced seismicity in enhanced geothermal systems using machine learning: challenges and opportunities”. In this study, Sadegh tests whether machine learning can forecast the cumulative seismic moment released during the next step of fluid injection. Such a forecast would allow operators to check whether the reservoir response remains within the stable range predicted by maximum magnitude models such as McGarr (2014) and Galis et al. (2017). He compared three models: a plain LSTM network, an LSTM combined with an attention layer, and the same architecture supplied with additional catalogue and operational features (b-value, correlation integral, seismogenic index, seismic efficiency ratio, clustered events ratio, and correlation with the McGarr model). The models were tested on induced seismicity from the Cooper Basin (Australia), the 2018 St1 Deep Heat stimulation in Helsinki (Finland), and a laboratory fluid injection experiment on a sandstone sample. The feature-rich model performed best in the Cooper Basin and laboratory cases, where seismicity was clustered and localized along a fault. For St1 Helsinki, with distributed and weakly interacting seismicity, the additional features brought little improvement. None of the models could anticipate the sudden jumps in seismic moment caused by the largest events. The paper discusses the main obstacles, such as limited training data from a single operational well and the need to extrapolate a cumulative quantity, and points to synthetic training data from numerical simulations as a way forward.
Reference:
Karimpouli, S., G. Kwiatek, P. Martínez-Garzón, D. Caus, L. Wang, G. Dresen, and M. Bohnhoff (2025). Forecasting induced seismicity in enhanced geothermal systems using machine learning: challenges and opportunities. Geophysical Journal International 242(2), ggaf155, doi 10.1093/gji/ggaf155