Automatic seismic phase picking using machine learning for the EGS Collab project

Arrival times of seismic phase(s) are necessary for both earthquake (small or large) location and subsurface structure imaging. Many algorithms have been developed to pick seismic phases automatically from seismograms. However, the accuracy of these automatic pickers is usually not satisfactory, especially for noisy data. Refined manual picks are still needed for reliable results. Results from recent applications of machine learning have shown that machine learning models can match or surpass human analysts on this subject. Many researchers have expressed their doubts on the generalization (transfer learning) of these machine learning models or techniques. We will show a generalization case of phase picking using seismic data recorded on a 3D subsurface array for the EGS Collab project. The sensors were installed in six boreholes at around 1.5 kilometers depth from the surface. The seismic array consists of 36 three-component sensors recording continuously at a sampling rate of 100,000 Hz. The seismic array collects nine trillion data points per day. The huge amount of data overwhelmed the analysts who pick seismic phases manually. On the other hand, accurate phase picks are required for both micro-seismicity monitoring and subsurface imaging. To attack this issue, we directly applied a machine learning model that was trained for natural earthquakes at a reginal scale and using data recorded at mostly surface stations with a different sampling rate. The preliminary results are promising. The machine learning model performs much better than a traditional algorithm (AR picker). The seismic phase arrival time measured by the machine learning model matches well with manual picks most of the time. The machine learning model is able to detect additional seismic phases (mostly S-wave) that were missed by analysts. We will use waveforms and labels gathered from the 3D seismic array to update the machine learning model and improve the performance.