A workflow for synthetic DAS data generation

Abstract

Distributed Acoustic Sensing (DAS) is an emerging data acquisition technology that utilises an optical fiber to measure dynamic strain along its axis. Composed by an optical fiber and an interrogator unit (IU), the system emits laser pulses into the fiber and detects phase shifts in the backscattered light, converting them into strain or strain rate measurements. DAS is becoming popular in many seismological applications and, in particular, for logistically challenging environments such as offshore areas, boreholes, glaciers, and volcanic settings, where deploying conventional monitoring is challenging. Spatial and temporal sampling of DAS systems is much higher than traditional seismological instruments, offering a detailed picture of the recorded seismic wave field. This high spatial and temporal sampling of DAS systems results in massive data generation, especially over extended acquisition periods. For instance, a single day's data collected with a 1 km fiber, featuring inter-channel distances of approximately 1m and a temporal sampling rate of 0.5 ms, can easily reach 2 TB. This highlights the need for efficient data analysis procedures in Distributed Acoustic Sensing (DAS) with methods that are both computationally fast and capable of exploiting the extensive information embedded in such data. As DAS data acquisition experiments are still few in numbers, generating and using synthetic data becomes essential for evaluating performance across diverse DAS acquisition geometries and testing new data analysis techniques. Despite the constant growth of DAS systems, there is a lack of standard modelling and analysis tools that can be used within routine procedures. To address this issues, we formulated a versatile workflow designed to generate synthetic DAS data based on the convolutional model. A central component of this workflow is a travel-time calculator based on the solution of the Eikonal equation, accommodating various data acquisition geometries, including scenarios involving optical fibers deployed in deep boreholes—whether vertical or oblique. Synthetic DAS seismograms are subsequently generated by using the computed travel times, for both P and S phases, with the convolutional model. These seismograms contain several information, such as the radiation pattern of the source and the directivity of the fiber, with the possibility of selecting an arbitrary wavelet. While DAS synthetics computed using the convolutional model may be less realistic than those generated with methods like the reflectivity or the spectral element method, their computational speed is much higher. This efficiency becomes particularly crucial when dealing with the generation of extensive DAS synthetic datasets. The synthetic generation workflow can be used for 1) testing new seismic event detection and location methods for DAS data and 2) training machine learning models. Lastly, this work includes a comparative analysis of synthetics obtained through our workflow against those generated using the spectral element method, followed by an application with a waveform-based DAS event detector.