ARGUS: A 17-ms End-to-End Deep Learning Pipeline for Real-Time Seismic Source Characterization and Ground Motion Prediction in Sparse-Network EGS/CCS Environments

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:Novel integration of GNN, spectral graph networks, and Fourier neural operators for real-time seismic source characterization; provides a benchmark for low-latency ML in geophysical monitoring.

🏢実務担当者:Enables cost-effective, real-time induced seismicity monitoring for CCS and EGS operations, reducing the need for dense sensor arrays and high-performance computing.

🏛政策担当者:Supports evidence-based regulation of induced seismicity for CCS and geothermal projects, demonstrating that accurate monitoring is feasible with sparse networks.

We present ARGUS (Automated Real-time Geophysical Understanding System), an end-to-end deep-learning pipeline that jointly estimates hypocenter location, centroid moment tensor (CMT), and peak ground acceleration (PGA) distribution from sparse seismic networks, targeting induced-seismicity monitoring in enhanced geothermal systems (EGS) and carbon capture and storage (CCS). From as few as four to eight stations, ARGUS produces all three outputs in 17.0 ms on commodity hardware. The pipeline chains three neural components: GNN-Locator, a graph-attention (GATv2) locator with conformal uncertainty quantification; SWIFT CMT, a spectral graph network for fracture-mechanism classification; and FNO-NAMI, a Fourier neural operator predicting 128 × 128 PGA maps in 4.5 ms. Because publicly available EGS microseismicity below Mw 2.0 was not used here, we validate on real regional records (K-NET and Hi-net, 2016 Kumamoto sequence) as a proxy spanning the EGS-relevant range Mw 2.6–4.0. The locator attains a median error of 10.3 km on random splits and 14.7 km on temporal splits after fine-tuning (12.7 km for Mw 2.6–4.0), with conformal intervals reaching 96.2% empirical coverage at the 90% nominal level; SWIFT CMT, trained on synthetic Utah FORGE data (99.4% three-class accuracy), transfers to 95.1% shear classification on Kumamoto, consistent with the documented strike-slip mechanism; FNO-NAMI reproduces PGA attenuation (Pearson r = 0.619; n = 2,892). We emphasize that this regional-scale validation establishes the feasibility of the integrated low-latency architecture rather than reservoir-scale EGS location accuracy, which requires the dedicated Utah FORGE field validation we outline as the immediate next step. Ablations confirm the contributions of GATv2 attention (+77%), S–P differential features (+53%), and the waveform encoder (+59%). Built entirely from public data on commodity hardware, ARGUS shows that simultaneous, sub-100-ms source characterization is attainable without high-performance computing.

• EarthArXiv https://eartharxiv.org/repository/object/13744/download/24117/first seen 2026-07-01 04:18:26