Detection of microseismic events in continuous DAS data using convolutional neural networks

Abstract

Detection and localization of microseismicity is an inevitable task in monitoring fluid injections into subsurface rocks during hydraulic stimulations. Traditionally, downhole geophones placed into boreholes or large surface geophone networks have been used for this, but in recent years, the use of distributed acoustic sensing (DAS) via fiber-optic cables placed into boreholes has become a common technique. However, DAS registrations still have lower signal-to-noise ratios than geophones; i.e., they cannot detect small-magnitude events. In this work, we develop and train a convolutional neural network capable of detecting microseismic events in continuous DAS recordings incorporating arrival-time information from geophones. The network is trained on DAS and geophone data from the Utah FORGE enhanced geothermal system project for which we are able to significantly shift the detection threshold toward smaller magnitude events. Although the number of microseismic events (approximately 150) used for training is small, the tested network performance is high and provides a complete event catalog down to magnitude MW = −1.6, a notable improvement over previous studies. Using a short recording period of several hours for training, such a network might be used for long-term, real-time monitoring of geothermal sites. Although the network is explicitly trained for the geometry of the data set used, the philosophy and network architecture can be adapted for similar case studies where long-term seismic monitoring is required (e.g., CO2 sequestration).