QuakeCastNet: an Interpretable Deep Learning Framework for Induced Seismicity Forecasting in Geothermal Fields

Abstract

Induced seismicity presents a critical challenge in geothermal reservoir management, as the occurrence of large seismic events raises public safety concerns and affects social acceptance of geothermal operations. Accurate forecasting of induced seismicity provides valuable information for operators and improves understanding of the underlying physical processes. However, physics-based modeling requires detailed knowledge of subsurface properties and intensive computation, while conventional statistical methods struggle to capture the nonlinear relationships between operational parameters and seismicity. In this study, we introduce QuakeCastNet, an interpretable deep learning framework for forecasting induced seismicity, including seismicity rates, the spatiotemporal evolution, and magnitude distribution in geothermal fields, and demonstrate it using data recorded at Utah FORGE and The Geysers. QuakeCastNet combines a modified Temporal Fusion Transformer (TFT) for time-series forecasting with a Graph Neural Network (GNN) for spatial representation learning, enabling joint modeling of complex dependencies across both time and space. The framework integrates heterogeneous datasets—including geological information, historical seismicity catalogs, and injection and production metadata—to predict future seismicity rates and magnitude-frequency distributions within gridded regions of interest. The estimated magnitude distribution provides insights into the probability of larger seismic events in the future to support risk assessment. By coupling explainable AI with data-driven learning, QuakeCastNet demonstrates improved predictive performance and interpretability, advancing our understanding of induced seismicity mechanisms and supporting safer, adaptive reservoir management strategies in geothermal operations.