Field VSP Lookahead Inversion for Velocity via Temporal Convolutional Networks

The inversion of zero-offset Vertical Seismic Profiling (ZVSP) corridor stacks for velocity and depth prediction is crucial for hazard prediction ahead of the drill bit, and for extending velocity information below the final depth of the well. Traditional approaches face challenges due to the non-linearity of inversion, limited data, and biases from wavelet choice, noise level, and other parameters. This study proposes a deep learning approach using VSP data only as input for the inversion, namely the measured velocity and the corridor stack. The model is trained and validated on 1D synthetics, then it was blind tested on additional 1D synthetics, 2D finite-difference modeling synthetics, and real datasets from Volve Field, Norway. The model predictions outperform conventional Lookahead VSP inversion approaches.