Speaker: Sydney Dybing (University of Oregon)

Abstract: Organizations such as the U.S. Geological Survey (USGS) and others need to be able to rapidly respond to and provide information about worldwide earthquakes across the magnitude scale in an accurate and timely manner, which often necessitates the use of a variety of different data types. Machine learning (ML) is a very useful tool for solving problems across the wide range of real-time earthquake response challenges. First I will discuss advancements in single-station earthquake magnitude estimation capabilities using an ML algorithm trained with a global dataset. However, given that magnitude estimates derived from seismic data for large earthquakes suffer from magnitude saturation due to the inherent limitations of inertial sensors, I will discuss how Global Navigation Satellite Systems (GNSS) data can be used to fill this gap. ML methods have been used to develop models capable of real-time tracking of large earthquake evolution with GNSS data and forecasting of ground motions to provide early warnings. We have also built an ML-based earthquake detector for noisy GNSS data, which if used in conjunction with the previous model could help reduce the risk of outlier predictions of source characteristics. Another approach to resolving this noise issue involves applying ML to de-noise GNSS data. Given the wide variety of scales, locations, and data types needed for earthquake early warning and rapid response, there are many promising applications of ML as a tool for solving these complex problems.