← Back to Sessions

Managing Induced Earthquake Potential with Deep Graph Neural Networks

Description: 

Earthquakes of societal and regulatory concern continue to occur in the unconventional oil fields of the U.S. In 2024 alone, 16 magnitude 4+ earthquake including 3 larger than magnitude 5 occurred in Oklahoma, Texas and New Mexico. Most of these events can be linked to deep injection of oilfield wastewater. Also of note was a M_W 4.6 event in the Eagle Ford of Texas, now the largest frac-induced earthquake in North America. As problematic seismicity continues to occur despite attempts to manage it through reductions in deep disposal, industry and regulators alike have expressed the need for better methods to identify areas with increasing potential for large magnitude earthquakes with sufficient lead time to take actions to mitigate the hazard.

In our study, we focused on employing deep learning techniques to explore the complex, non-linear relationships between injection activities and earthquake hazards in Oklahoma. Our primary aim during model development has been to develop a model that incorporated disposal/production data that outperforms baseline earthquake hazard estimators based solely on seismicity rate. We developed an attention-based graph neural network (GNN) model “InFormer” that utilizes two attention mechanisms: one temporal, and one spatial. This architecture allows encoding of both long-term and long-distance context to capture far-reaching relationships. A simple rolling mean of seismic activity is used as a baseline for comparison with training results. Results are encouraging for models trained with a 6-month context and a 3-month forecast horizon. These initial experiments illustrate the potential of deep learning to forecast evolving earthquake hazard by leveraging temporal injection histories. By incorporating deep nonlinearities on the feature set, InFormer demonstrates strong temporal generalization, achieving 30% to 40% improvement in model validation over 3-month and 6-month rolling mean baselines. Transfer learning continues to be challenging, as the seismicity response to a given volume of injected wastewater in different areas may deviate significantly.

Session: From Physics to Forecasts: Advancements and Future Directions of Induced Seismicity Research - I

Type: Oral

Date: 4/15/2025

Presentation Time: 08:30 AM (local time)

Presenting Author: Brandon

Student Presenter: No

Invited Presentation: 

Poster Number:

Authors