Attn-Loc: Deep Learning for Microseismic Event Location via Transformer Encoder

Abstract

Microseismic event location is a crucial geophysical monitoring task in engineering applications such as hydraulic fracturing, geothermal development, carbon dioxide geological storage, and mining safety monitoring. Traditional location methods rely on travel-time differences and velocity models, making them sensitive to model errors and observational noise and difficult to adapt to varying observation geometries. In recent years, deep learning–based methods have shown potential for end-to-end location in noisy environments but are often limited to fixed station layouts. This article proposes Attn-Loc, a microseismic location method based on the Transformer encoder. By inputting receiver coordinates, P- and S-wave arrival times, incidence angles, or azimuth angles, it constructs a probabilistic source coordinate estimation model that effectively integrates multisource uncertainties inherent in observational data. The method utilizes the self-attention mechanism to implicitly learn interstation correlation features and embeds regional velocity structures into the network parameters, enabling flexible adaptation to various observation geometries, including surface, downhole, and time-lapse monitoring. We employ compute unified device architecture (CUDA)-accelerated raytracing to generate large-scale training samples and train lightweight Transformer encoder–based networks, validating its high accuracy and strong robustness on both synthetic and field data. Through localization error assessment, Attn-Loc achieves small multidirectional relative errors on synthetic data and exhibits high consistency with existing catalogs in field data from the Tony Creek Dual Microseismic Experiment and Frontier Observatory for Research in Geothermal Energy projects, confirming its practicality and versatility in complex geological environments and under variable observation geometries.