Том: 15406 LNCS , Год издания: 2024

High-resolution subsurface monitoring is a prerequisite for reliable seismic imaging of geological structures. Seismic Full Waveform Inversion (FWI) has emerged as a promising technique in exploration geophysics to investigate such media. However, these kinds of studies need extremely time- and memory-consuming computations, which prevent the FWI from being used in practice to estimate the subsurface changes in real-time, especially when evaluating the viscoelastic properties of target geological objects. This paper introduces the supervised end-to-end data-driven inversion approach to establish a nonlinear mapping between changes in raw seismic data and corresponding changes in subsurface properties for viscoelastic media, such as P-wave velocity (vp later) and corresponding quality factor of P-wave (Q_p factor later). In the proposed approach, significant computational resources are consumed during the generation of a training dataset and training of the neural network, which typically requires more time than conventional FWI. However, once the network is trained, the prediction stage dramatically reduces the computational cost required for inversion once a good generalization capability of the neural network is achieved. To eliminate these limitations, the Born approximation and employed transfer learning are demonstrated for accelerating the generation of the training dataset, provided that the changes in the viscoelastic model of the medium are slight. The performance of the data-driven approach with Born approximation was evaluated using synthetic seismic data generated by a 2D frequency viscoelastic solver for the Gullfaks model deposit.