Improving Real-Time Forecasts of Induced Seismicity Through Machine Learning-Based Event Classification
Avigyan Chaterjee
Lawrence Livermore National Laboratory
- Date & Time
- Location
- Hybrid In-Person and Online via Microsoft Teams
- Host
- Tara Nye
- Summary
Accurate and rapid classification of seismic events is essential for real-time monitoring and informed decision-making in subsurface industrial operations. In this project, we advance seismic event classification by developing a robust machine learning (ML) pipeline designed to distinguish events of interest (such as local earthquakes) from extraneous signals, including anthropogenic noise, weather-related disturbances, and teleseismic events, in real time. Utilizing a labeled dataset of over 100,000 seismic events and false positive detections, we assess several deep learning architectures, including Convolutional Neural Networks (CNN), U-Net, and U-Net models enhanced with self-attention mechanisms. After initial binary classification at the single-station level, predictions from all stations are combined to produce a final event classification across the network. Our experimental results show that the top-performing model achieves a classification accuracy of 95% and an area under the Receiver Operating Characteristic curve (AUC) of 0.95, substantially reducing false positives compared to traditional methods. Integrating this ML-based classification module into the real-time seismic processing workflow is expected to minimize manual review by filtering out likely false detections. Ultimately, this approach enables the delivery of high-fidelity, real-time seismic catalogs to industrial operators and the DOE’s ORION toolkit, supporting improved operational forecasting and hazard assessment.