PINNHYPO: HYPOCENTER LOCALIZATION USING PHYSICS INFORMED NEURAL NETWORKS

Many industrial activities needed to sustain human society have the potential to induce earthquakes. With the increasing availability of data and computational resources, researchers have started to exploit the capabilities of machine learning algorithms to detect, locate, and interpret seismic events. For hypocenter localization, typically a convolutional neural network (CNN) is trained in a supervised manner using a historical or synthetically generated dataset. However, this approach often requires a huge amount of labeled data that may not be readily available. Therefore, we propose a hypocenter location method based on the emerging paradigm of physics-informed neural networks (PINNs). Using observed P-wave arrival times for an event, we train a neural network by minimizing a loss function given by the misfit of observed and predicted traveltimes, and the residual of the eikonal equation. The hypocenter location is then obtained by finding the location of the minimum traveltime in the computational domain. Through synthetic tests, we show the efficacy of the proposed method in obtaining robust hypocenter locations, even in the presence of sparse traveltime observations. This is due to the use of the eikonal residual term in the loss function that acts as a physics-informed regularizer.