Inversion of Vertical Seismic Profiling (VSP) Data via Temporal Convolutional Networks

Abstract

Inversion of Vertical Seismic Profiling (VSP) data for velocity prediction ahead of the drill bit is crucial for hazard prediction but challenging using traditional methods, which are sensitive to data-dependent parameterization and physical assumptions. These traditional workflows are often limited by key assumptions, such as the sparsity of the Earth’s reflectivity series and the presence of a time-invariant wavelet, which make the inversion problem inherently non-linear and non-unique. To address these limitations, this study proposes a deep learning solution, introducing a dual-input neural network that inverts for velocity using two distinct inputs: the VSP corridor stack, which contains reflections from above and below the bit, and the measured velocity profile from first breaks, which constrains the drilled section. For comparison, a single-input network that uses only the corridor stack was also evaluated against the traditional Lookahead inversion method. The networks were trained on a vast and diverse set of 1D synthetics to promote generalization and then blind-tested on 1D and 2D finite-difference synthetics, as well as real VSP datasets from the Volve Field, Norway. The dual-input network significantly outperformed both the traditional method and the single-input network, demonstrating better accuracy and robustness. The single-input network performed on par with the traditional approach, highlighting the critical advantage of incorporating measured velocity as a constraint to better constrain the otherwise ill-posed inversion problem. The results demonstrate that the proposed dual-input deep learning network provides a more systematic and reliable approach for Lookahead VSP inversion, relying solely on data from the VSP survey itself.