Using Machine Learning Techniques with Incomplete Polarity Datasets to Improve Earthquake Focal Mechanism Determination

Abstract Earthquake focal mechanisms are traditionally produced using P-wave first-motion polarities and commonly require well-recorded seismicity. A recent approach that is less dependent on high signal-to-noise exploits similar waveforms to produce relative polarity measurements between earthquake pairs. Utilizing these relative polarity measurements, it is possible to produce composite focal mechanisms for clusters within microseismic sequences using regional networks. However, missing or low-confidence polarity measurements still limit our ability to calculate high-quality composite focal mechanisms. Here, we replaced unreliable polarity measurements with estimates using iterative random forests, an unsupervised ensemble machine learning method. Using the imputed (“replaced”) polarity data, we then categorically clustered the events into families. As a case study, we applied this modified composite mechanism workflow to a multistation template matched catalog of an earthquake swarm that occurred during 2020 near the Maacama fault in northern California. We found that our modified methodology produced higher-quality earthquake families and improved composite focal mechanisms, with fault-plane uncertainties <35° for 94% of the families compared with 34% of families using the previous methodology.