QuakeCast: Forecasting Earthquakes from Preseismic Ionospheric Signals Using Machine Learning Refinements and Advances

QuakeCast is an earthquake forecasting system that aims to provide risk assessments in the 8 hours prior to major earthquake events for regions affected by earthquakes. Earthquakes are one of the worlds most fatal and expensive natural disasters, causing over 2.2 million fatalities and over $800 billion in damages within the last century. Current earthquake early warning systems provide up to one minute of warning by detecting an earthquakes faster, non-damaging P-waves before the arrival of the slower, damaging S-waves. However, if earlier warning times could be achieved, critical preparatory activities, such as staging resources, securing critical infrastructure, and even evacuation could greatly reduce fatalities and damage costs globally. Electromagnetic ionospheric phenomena have been observed days before major earthquakes, notably in 2015 in Nepal. One theory is that microcracking in the lithosphere prior to the crust rupture event generates an electric current that interacts with the ionosphere through the global electric circuit. At AGU 2020, we presented our initial findings for using machine learning to detect these signals within Total Electron Content (TEC) data in the ionosphere. We compiled a global dataset of ionosphere and earthquake data from 2005-2015, and developed a pair of machine learning classifiers a classical logistic regression model, and a deep learning ConvLSTM autoencoder anomaly detector on the dataset. Both classifiers were able to predict preseismic signals significantly better than chance, and up to 8 hours in advance. This year, we address some of the limitations of our previous approach. First, we present techniques for inferring the location of the earthquake within the sequence window from last years results. Second, we present an updated dataset with additional years of ionosphere and earthquake data, ionosphere dynamics information, ionosphere-modeling data, and a high-resolution dataset that captures mesoscale features in the ionosphere. Finally, we discuss and present preliminary results from new models in development, which seek to improve accuracy from 65% to our goal of 85%, as well as enable direct prediction of location, time, and magnitude.