Forecasting the spatiotemporal evolution of fluid-induced microearthquakes with deep learning

Microearthquakes (MEQs) produced by subsurface fluid injection record the spatial and temporal evolution of stress and permeability. Hence, the ability to forecast the spatiotemporal evolution of fluid-induced microseimicity is crucial for guiding applications such as Enhanced Geothermal Systems (EGS) and CO2 sequestration. We present a transformer-based deep learning approach for predicting the spatiotemporal evolution of MEQs. The model is trained on hydraulic stimulation history and monitored MEQ events from the EGS Collab field experiment data. Our transformer network accurately forecasts MEQ occurrences, energy release, and spatial extent, even for unseen datasets. Notably, the predicted spatial extents after the fluid injection period were within 0.36 meters of observed values, demonstrating the model's robustness in capturing the spatial characteristics of fluid-induced seismicity. These results underscore the significant potential of deep learning approaches to accurately forecast MEQs, ultimately improving risk assessment, guiding mitigation strategies, and optimizing well placement in future geo-engineering operations.