← Back to Sessions

Machine Learning Phase Picking and Association Reveals a More Complete Induced Earthquake Catalog for the Fort Worth Basin, Tx

Description: 

The North Texas Earthquake Study (NTXES) in the Fort Worth Basin, USA, comprised a heterogeneous network of short-period, strong-motion, broadband, and nodal sensors spanning a large metropolitan area. It was deployed to capture earthquakes associated with oil and gas extraction, primarily wastewater disposal operations. The evolving sensor configurations from 2013-2025, combining velocity and accelerometer data, pose challenges for metadata processing in machine learning (ML) applications, as few training datasets incorporate accelerometer data. This study utilizes SeisBench to implement PhaseNet and EQTransformer (Zhu & Beroza, 2019; Mousavi et al., 2020) for phase picking, and explores traditional and clustering association techniques to decrease the magnitude of completeness of the published NTXES catalog. The original PhaseNet training dataset from Northern California provides the best performance for our induced seismicity detection, as it includes data from different seismic instrument types similar to our heterogeneous network, though some stations require optimized training datasets. Comparing automated detections against the manually curated catalog, only 55.7% (1,238 events) are matched; however, our workflow identifies 4,617 additional events. For matched events, ML-based locations show good agreement with manual picks, with mean differences of 0.77s in origin time, 3.4 km in latitude, 2.6 km in longitude, and 1.8 km in depth. Statistical validation across stations IPD1 (Irving, urban short-period), AZDA (Azle, rural short-period), and VTAX (Venus, NetQuakes accelerometer) reveals high precision in automated phase picking. P-wave arrival time mean residuals are 0.0077s (IPD1), 0.0393s (AZDA), and 0.0045s (VTAX), with S-wave residuals of -0.0385s, 0.0792s, and 0.0363s respectively. S-P time differences show excellent agreement (means: 0.0003s, 0.1097s, 0.0617s). We show that PhaseNet paired with its original training dataset significantly enhances earthquake detection in induced seismicity monitoring networks, providing more complete catalogs essential for hazard assessment.

Session: From Drilling to Ground Shaking: Mechanisms, Monitoring and Mitigation of Induced Earthquakes [Poster]

Type: Poster

Date: 4/17/2026

Presentation Time: 08:00 AM (local time)

Presenting Author: Yeshey Seldon

Student Presenter: Yes

Invited Presentation: 

Poster Number: 101

Authors