AN EARTHQUAKE LOCALIZATION WORKFLOW BASED ON HIERARCHICAL CONVOLUTIONAL AUTOENCODERS

The localization of seismic events is one of the fundamental problems in seismological research. The expanding seismic network has led to a rapid increase in the demand for localization tasks; however, traditional nonlinear Bayesian earthquake localization methods based on numerical solvers of the Eikonal equation are often difficult to parallelize and require high computational costs, especially for fine-grained localization due to the use of dense grids. Our work proposes a workflow primarily based on convolutional autoencoders for estimating the fine-grained probability distribution of epicentral locations using the observed P and S arrival times from seismic stations. Firstly, a coarse-grained, low-resolution source probability distribution is simply computed from the observed arrival times at the stations. Then, we utilize a set of hierarchical convolutional autoencoders to progressively expand this distribution. Eventually, we obtain a fine-grained, high-resolution probability density distribution. These autoencoders are trained incrementally using the kernel density estimations (and their down-samplings) of the samples provided by the NonLinLoc program. The final high-resolution probability density distribution not only provided the mean or maximum-likelihood predicted seismic location but also the uncertainties of the solution. We have validated the proposed workflow through synthetic seismic event localization experiments using various synthetic velocity models and a real velocity model from mainland China. The experiments demonstrate that appropriately trained networks can achieve comparable accuracy to traditional tools (such as NonLinLoc) , while significantly enhancing the computational efficiency.