Work

Earthquake location and magnitude estimation using seismic arrival times pattern and gradient boosted decision trees

Year: 2025

Type: article

Abstract: We present a machine learning methodology designed for the accurate determination of seismic source parameters. These parameters include earthquake location, origin time, magnitude, and associated uncertainties such as horizontal and depth errors, and azimuthal gap. Our approach capitalizes on temporal patterns derived from earthquakes within a seismic network and employs the Histogram-Based Gradient Boosting method. This method is recognized for its superior prediction accuracy and computational efficiency compared to other Decision Tree algorithms. To demonstrate the robustness of our proposed methodology, we generated a synthetic earthquake bulletin. This bulletin mirrors standard seismic network configurations, station-event contributions, and implements a complex 3D velocity model for travel time ray-tracing. We included varying amounts of P and S phases, as well as amplitudes for each simulated earthquake. The temporal patterns recorded from these events served as input data, with the earthquake location, magnitude, origin time, horizontal error, depth error, and azimuthal gap used as targets. We created three datasets based on data completeness and used $R^2$ score, Mean and Median Absolute Error (MAE and MEDAE) to thoroughly compare predicted and true values for various location attributes. The results highlight the model's proficiency in accurately estimating event location and time, and show that the predicted magnitudes closely match the true ones. Furthermore, statistical parameters such as horizontal error, depth error, and azimuthal gap display very small MAE/MEDAE, indicating high accuracy in determining the uncertainty of the events' location. A notable advantage of the proposed method is its high execution speed, especially when compared to routine location programs. Furthermore, its ability to inherit location approaches from various earthquake location programs is a notable strength. Given these characteristics, it could be effectively implemented in permanent seismic networks and early warning systems, where a short execution time combined with high accuracy is vital. We conducted further tests on the estimator model using the Ahar-Varzaghan 2012 doublet earthquake aftershocks sequence, recorded by a local seismic network in NW Iran, as a real-world case study. The results confirm the robustness and effectiveness of our model, even in the presence of significant missing data. Additionally, the results, both in map view and cross sections, are consistent with previous high-resolution seismicity studies.

Source: Artificial Intelligence in Geosciences

DOI: 10.1016/j.aiig.2025.100149

Authors: Saeed SoltaniMoghadam, Anooshiravan Ansari, Leila Etemadsaeed, Mohammad Tatar, Meysam Mahmoodabadi

Institution: International Institute of Earthquake Engineering and Seismology

Language: English

Topic: Seismology and Earthquake Studies

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $700

Funder: International Institute of Earthquake Engineering and Seismology