Joint Microseismic Event Detection and Location with a Detection Transformer

TL;DR

The paper tackles the real-time joint detection and localization of microseismic events by adapting the DEtection TRansformer to seismic waveforms. It introduces a set-based prediction framework with bipartite matching and a Hungarian loss, trained on synthetic data and enhanced by MLReal domain adaptation to perform robustly on field data. Validation on a SEAM Time Lapse 2D model and Arkoma Basin field recordings shows high classification and location accuracy with millisecond-scale inference, outperforming traditional migration in speed and enabling real-time monitoring. The approach promises a scalable path toward end-to-end, real-time microseismic monitoring, with clear routes for extension to 3D, elastic waves, and more complex field conditions, albeit with reliance on a reasonable velocity model and robust data preprocessing.

Abstract

Microseismic event detection and location are two primary components in microseismic monitoring, which offers us invaluable insights into the subsurface during reservoir stimulation and evolution. Conventional approaches for event detection and location often suffer from manual intervention and/or heavy computation, while current machine learning-assisted approaches typically address detection and location separately; such limitations hinder the potential for real-time microseismic monitoring. We propose an approach to unify event detection and source location into a single framework by adapting a Convolutional Neural Network backbone and an encoder-decoder Transformer with a set-based Hungarian loss, which is applied directly to recorded waveforms. The proposed network is trained on synthetic data simulating multiple microseismic events corresponding to random source locations in the area of suspected microseismic activities. A synthetic test on a 2D profile of the SEAM Time Lapse model illustrates the capability of the proposed method in detecting the events properly and locating them in the subsurface accurately; while, a field test using the Arkoma Basin data further proves its practicability, efficiency, and its potential in paving the way for real-time monitoring of microseismic events.

Figures (12)

• Figure 1: The architecture of DEtection TRansformer. The five output slots indicate that a maximum of five events can be detected and located per segment. In practice, this number should be set to a value significantly larger than the expected number of events within a single seismic segment to ensure that all potential events can be captured.
• Figure 2: The diagram of Feed Forward Network branches. Fed with the prediction sequences from the Transformer, the classification branch produces the probability predictions $\hat{p}$ of event existence, while the regression branch produces the coordinate predictions $\hat{s}$ of corresponding source locations.
• Figure 3: The MLReal transformation diagram for (a) training data and (b) application data, respectively.
• Figure 4: Illustration of two common scenarios where the predicted number of events does not match the ground truth. In case (a), the input segment contains two events, but the network only detects one event and misses the second. In case (b), the input segment contains a single event, but the network over-predicts to two events. The colored arrows indicate the optimal assignment determined by the Hungarian algorithm.
• Figure 5: The SEAM Time Lapse model. The smaller red box represents the expected area of microseismicity and is used for generating test (assumed real) data, while the bigger black box is selected and used for generating training data. The receivers, denoted with yellow dots, are deployed on the surface at a regular interval of 30 meters.