Challenges and Opportunities of Machine Learning Earthquake Detection for Regional Monitoring

ABSTRACT Machine learning (ML) earthquake phase detection algorithms continue to gain popularity and are routinely used to generate research catalogs with thousands of previously uncatalogued events. Many ML algorithms are available pretrained on large, global data sets. These pretrained models promise regional transferability and applications in real-time monitoring organizations. However, the adoption of these ML algorithms by monitoring agencies requires trusted performance across a wide range of seismic monitoring challenges. We apply a pretrained algorithm to four characteristic studies representing a range of network, tectonic, and environmental challenges. We establish a three-catalog comparison framework between our ML catalog, a real-time catalog, and an analyst reviewed catalog. We visually assess and label all ML and real-time events not found in our analyst reviewed catalog. Finally, we subset all additional events that match our catalog standards, establishing a one-to-one performance comparison between our ML and real-time algorithmic catalogs. For each study, we find our ML catalog provides a consistently higher match to our analyst reviewed catalog than the real-time catalog. However, the ML catalog from each study introduces additional complexities ranging from the large addition of poorly constrained, smaller magnitude events; misidentification of nonearthquake signals; and missed detection of large magnitude, felt earthquakes. These discrepancies warrant further training data set scrutiny and suggest that the establishment of a location-based training data set is necessary for consistent and reliable ML performance.