Cycle-consistent convolutional neural network for seismic impedance inversion: An application for high-resolution characterization of turbidites reservoirs

Abstract

Acoustic impedance is a subsurface layer parameter crucial for reservoir characterization due to its relationship with petrophysical properties. Seismic impedance inversion is the routine conventionally used to calculate the acoustic impedance in 3D seismic datasets. Deep learning-based seismic inversion has recently gained attention due to its capacity to establish non-linear relationships between observed data and model parameters, producing robust acoustic impedance estimates. We employed a 1D cycle-consistent convolutional neural network (CNN) to perform the high-resolution seismic impedance inversion in the turbidite reservoirs of the Jubarte Field, Campos Basin, Brazil. The neural network was trained using geostatistics-based pseudo-wells with high pattern variability. Before applying the trained CNN, we performed the seismic data pre-conditioning to remove high and low-frequency noises affecting the data amplitudes, making the dataset more suitable for seismic impedance inversion. Our results show that the deep learning-based inversion produced a high-resolution estimate, allowing an accurate internal characterization of turbidite lobes. Quantitatively, the estimated average correlation coefficient in the eight wells evaluated in this study was 0.78. We observed that the pre-conditioning step was important for this application since the 1D architecture utilized could not deal properly with the noise as it disregards lateral connections. 2D and 3D networks may address this issue. Compared to the open-loop CNN and the traditional model-based inversion, the cycle-consistent network produced the best estimate, with good lateral continuity, vertical resolution, and correlation. We support that modern deep-learning architectures like the one presented can be efficiently integrated into reservoir characterization workflows for enhancing subsurface assessment.