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

We present a machine learning approach for earthquake location and magnitude estimation based on seismic arrival time patterns, using Histogram-Based Gradient Boosting for its high accuracy and computational efficiency. The model is first evaluated using a synthetic earthquake bulletin that simulates realistic network geometry, station-event distributions, and incorporates a 3D velocity model for accurate travel-time computation. Input features include P and S arrival times and amplitudes, while targets consist of location, origin time, magnitude, and uncertainty measures (horizontal and depth errors, azimuthal gap). Model performance is evaluated using $R^2$, Mean Absolute Error (MAE), and Median Absolute Error (MEDAE), demonstrating high accuracy across datasets with varying levels of completeness. Finally, we validate the model using real-world data from the Ahar-Varzaghan 2012 aftershock sequence in NW Iran. The model accurately recovers key spatial patterns of seismicity despite significant missing data, and the results align with previous high-resolution studies. These findings confirm that the proposed method generalizes well beyond synthetic settings and offers a fast, robust alternative for operational seismic networks and rapid hazard assessment.