RPNet: Robust P-Wave First-Motion Polarity Determination Using Deep Learning

Abstract We present RPNet, a robust deep-learning model for P-wave first-motion polarity determination aimed at deriving earthquake focal mechanism solutions. RPNet integrates advanced deep-learning techniques, including inception modules, attention mechanisms, and Monte Carlo dropout, to improve prediction accuracy and stability with quantified uncertainty. We conduct benchmark tests against four existing models using data sets from the western United States and Italy. The results address key issues such as sensitivity to misaligned P-wave onsets, class imbalance between up and down polarities, and handling of “unknown” labels. To enhance robustness, we apply intensive random time shifts of P-wave onsets within a ± 0.5 s range during training. To address the class imbalance, RPNet is trained with an equal amount of data for both up and down polarities. The performance of RPNet is evaluated through three tests. First, we test RPNet on a data set comprising 10% of the data excluded from training, achieving 99% recall for both up and down polarities. Next, we apply RPNet to independent Hi-net data from Japan, where it demonstrates superior generalization compared to previous deep-learning models, achieving recall rates above 97.5% even under time shifts of up to ±0.40 s. Finally, RPNet is tested on the 2016 Kumamoto earthquake sequence, where its automatically derived focal mechanism solutions closely match those manually derived by the Japan Meteorological Agency, outperforming the previous model in Kagan angle and polarity misfit. These results highlight RPNet’s potential as a reliable tool for automating focal mechanism derivation across diverse regions and waveform conditions without the need for additional optimization.