The Impact of Tectonic Setting on Machine Learning Approaches for Earthquake Prediction

In prior decades the concept of using mathematical methods to predict earthquakes was considered infeasible. Recent advances in machine learning and predictive modeling offer promising avenues to potentially realize earthquake prediction. In order to test the viability of machine learning methods, experiments were made with earthquake datasets from Kansas and Puerto Rico. The two datasets were chosen for the distinct differences in their tectonic settings. Kansas has few major faults, with a largely inactive subsurface, this produced a smaller dataset with a few large clusters. Puerto Rico is complexly faulted, with an extremely active tectonic setting, this produced a larger dataset with a large number of small clusters. In order to test the effectiveness of these two datasets for machine learning and prediction they were run through three different machine learning algorithms including an LSTM model, Bi-LSTM model, Bi-LSTM model with attention. Not only were the three different machine learning methods compared against each other for accuracy but also the datasets as well. Conclusive findings show that the two different data sets favor different processing methods. The Kansas data performs the best with the Bi-LSTM with attention model, while the Puerto Rico data performs the best with the LSTM model. This is likely due to the tectonic settings of the two regions, since the Kansas dataset has less overall data, and earthquakes are concentrated in a few large clusters, while the Puerto Rico data set has a more even distribution.