FocoNet: Transformer‐Based Focal‐Mechanism Determination

Abstract Traditional focal‐mechanism determination primarily relies on fitting the first‐motion polarities with grid‐search algorithms. We developed a machine‐learning model, FocoNet, to include more seismic information into focal mechanism determination. We designed FocoNet with transformer encoders, so that it learns how different stations relate to each other (each with seismic information encoded as a token vector). FocoNet combines first‐motion polarities and other seismic phase information (e.g., S/P amplitude ratios from different channels and signal‐to‐noise ratios) as its input. We evaluated FocoNet on both synthetic and real test cases. Compared with established grid‐search algorithms, FocoNet achieves higher focal mechanism accuracy and stability that manifests in 7–14° average improvements in Kagan angles relative to the ground truth labels. FocoNet's predictions achieve the greatest improvement in the challenging, but typical, situation of limited station coverage. We present two reasons why FocoNet outperforms traditional grid‐search algorithms. First, the input includes additional phase information that helps to constrain the solution. Second, the n