Deep Learning Driven DAS Strain Conversion to Geophone Ground Motion

Distributed Acoustic Sensing (DAS) has become a revolutionary observational technology for different geophysical applications. DAS, known for its high spatial resolution, environmental resilience, and ease of deployment, which make it a potential replacement to the traditional physical sensors that have been used for decades in seismology. The primary distinction between DAS and conventional seismic sensors lies in the fact that DAS inherently captures strain (or strain rate), in contrast to seismic instruments which record translational ground motions. However, the problem of strain directional sensitivity poses challenges for its direct use in standard seismic analysis. Therefore, several physics-based methods have been proposed to convert DAS strain to ground motion response (displacement, velocity, or acceleration). Efficient conversion of strain to ground motion using physics-based methods relies on accurate estimation of phase velocity along the DAS cable which is unavailable in most cases. To overcome this problem, we introduce a novel deep learning (DL) approach to convert high-resolution Distributed Acoustic Sensing (DAS) strain measurements into ground motion (GM).  The DL model employs a Bidirectional Long Short-Term Memory (BiLSTM) network. The model is trained and evaluated utilizing data from the PoroTomo project at Brady Hot Springs Geothermal Natural Laboratory. This dataset includes earthquake waveforms recorded by collocated DAS channels and geophones. The model’s performance is evaluated using the Root Mean Squared Error (RMSE) metric. It demonstrated an average RMSE of 0.41 for training and 0.95 for testing, indicating the model's efficacy in transforming DAS strain to particle velocity. The comparison results of predicted and original geophone waveforms further validated the model's accuracy within the relevant frequency range. This study marks a significant advancement in adapting high-resolution DAS strain data for use with conventional seismic data analysis techniques, thereby expanding the capabilities of seismic monitoring and interpretation.