Learning Injection–Seismicity Coupling for Probabilistic Multi-Horizon Forecasting in Geothermal Systems

Abstract Induced seismicity presents significant challenges to the sustainable development of geothermal energy, as fluid injection and reservoir stimulation frequently trigger seismic events. The underlying physical mechanisms involve complex interactions between the solid Earth and injected fluid, which makes the numerical flow modeling and mechanical failure analysis difficult, and leads to inaccurate forecasting of induced seismicity. This study introduces a deep learning framework that integrates multi-horizon forecasting with interpretability to predict seismicity in geothermal fields. By dynamically selecting relevant seismicity and operational data through variable selection networks, the model unifies key drivers within a single predictive architecture. The model with self-attention mechanisms and probabilistic forecasting identifies key seismicity drivers and quantifies uncertainty to evaluate fluid injection on future seismicity. Case studies from two geothermal fields, the Geysers and Utah FORGE, reveal site-specific seismic response governed by hydromechanical conditions, offering insights into fluid-induced stress evolution. Beyond risk mitigation, this approach provides a data-driven pathway for probing subsurface rheology through seismic response to geothermal operations.