Developing and Applying Deep Learning Time-Frequency Denoising Tools to DAS Data to Enhance Passive Seismic Signals for Seismic Hazard Analysis and Reservoir Characterization

DAS (distributed acoustic sensing) has been widespread in the fields of glaciers, volcanoes, fault zones, and geothermal, etc. due to its high spatial and temporal resolution, which also shows big potential to be used for seismic characterization and monitoring during carbon storage. However, generally the large DAS data highly contaminated by serious noise appear to have a very low signal-to-noise ratio, which prevents the wide application of DAS data. The team is developing an efficient and effective DAS denoising tool based on autoencoder deep learning neural networks by assembling and labeling borehole DAS seismic events from different sites. In Phase I, the team developed an effective cascade workflow for labeling field borehole DAS seismic events, and end-to-end autoencoder deep learning DAS denoising tool with high-efficiency targeting on borehole DAS data. The case study on the earthquake and microseismic events from FORGE site indicate that the developed deep learning DAS denoising tool is superior to the conventional industry spectral filtering method. In Phase II, the team will 1) further optimize the deep learning based DAS denoising tool by expanding more types (especially surface DAS array) of assembled and labeled DAS seismic events; 2) develop a workflow to monitor the induced seismicity during carbon storage using the denoised DAS seismic data; 3) Build carbon storage DAS array design tool for general use. The DAS denoising and seismicity monitoring tool is critical for the carbon storage site characterization and monitoring in Phase 3 of field case study. Additionally, the developed tool is naturally to be extended to other fields for DAS applications, such as seismology, geothermal, etc.