Relocating Acoustic Emission in Rocks with Unknown Velocity Structure with Machine Learning

Abstract

Inversion of hypocenters is the first and most fundamental step in the study of seismic activities. It requires solving the nonlinear relation between the travel time and hypocenter locations, which is heavily dependent on the knowledge of the medium properties, most importantly the velocity structure. In this study, we prove that machine learning (ML) methods including artificial neural networks (ANNs) and support vector machines (SVMs) can relocate hypocenters without a priori knowledge of the velocity structure. We train ML models with acoustic emissions (AEs) created by breaking pencil leads at known locations on a laboratory fault, using the relative P-wave arrival time as the input and AE source locations as the output. The resultant ML models can accurately relocate AEs on the fault surface. With carefully chosen training strategies, the ANN model achieved better accuracy than the SVM model. This study suggests that ML methods can provide effective and accurate approaches for relocating seismic events in a medium with unknown velocity structures.