DL FWI of dual-well DAS-on-tubing VSP data

Abstract Distributed acoustic sensing (DAS) is increasingly used for vertical seismic profiling (VSP) due to its high spatial sampling density and reduced costs relative to conventional geophone-based methods. This paper presents a large-scale dual-well DAS VSP survey and a comprehensive workflow for data conditioning and early-arrival full-waveform inversion (FWI) implemented using the deep learning (DL) framework PyTorch. Two adjacent wells, 4 km deep and 1.5 km apart, were instrumented with fiber on tubing. Dense land acquisition spanning 4.5 km along the line connecting the wells features a single dual-fiber interrogator and high levels of noise driven by the arid environment. We present a fully automated workflow that allows processing of such data introducing zero-offset supergather that is kinematically inverted for initial model building and synthetic-driven muting strategy to mitigate the effect of the noise. We also introduce pressure to strain-acceleration transform to accommodate DAS simulations using an acoustic solver and source function centroid constraints to adaptively update the source function throughout the iterations while preserving the source timing. Our results confirm the feasibility of combining DAS data and a DL framework for high-resolution velocity inversion in a desert environment with fiber deployed on tubing.