QuakeCastNet: Source Code for Forecasting of Injection-Induced Seismicity

This submission houses the companion code for the manuscript by Zhengfa Bi (Lawrence Berkeley National Laboratory) and Nori Nakata (Lawrence Berkeley National Laboratory; MIT). Both the companion code and manuscript are included in the resources section of this submission.

Fluid injection and reservoir stimulation at geothermal fields can trigger induced seismicity, and reliable forecasting is complicated by the coupled interactions between the solid Earth and the injected fluids. QuakeCastNet is an interpretable deep learning framework that issues probabilistic, multi-horizon forecasts of seismicity rates and their spatial distribution, combining historical seismicity with past and scheduled injection operational data. The framework consists of three components: Temporal branch: an extended Temporal Fusion Transformer (TFT) produces quantile forecasts of the seismicity rate, with multi-seed ensemble inference for stability; Post-injection decay adaptation: an empirical decay shape learned from historical shut-in episodes adapts the forecast once injection stops; Spatial branch: a regime-adaptive spatial density model (RA-SDM) blends a recency-weighted kernel estimate, a physical spreading kernel, and local triggering around larger historical events into a normalized density field, with the injection state selecting between stable and active regimes.

Case studies at two geothermal fields, The Geysers (California) and Utah FORGE (2024 stimulation), reveal site-specific seismic responses governed by hydromechanical conditions: the model infers a longer predictive memory at The Geysers and a more immediate relation to operations at Utah FORGE. Attention weights and input-variable ablations provide interpretability, and temporal and spatial forecasts are benchmarked against statistical reference models.