A pre-training and fine-tuning paradigm for building a subsurface model

Abstract Estimating the subsurface model from various geophysical measurements is a crucial task in hydrocarbon exploration and development. Seismic data has a wide coverage, but the resolution of the seismic full waveform inversion is constrained by the frequency band of the seismic data. Well loggings measure the medium parameters with high resolution but only have sparse coverage. We employ the Transformer architecture to integrate well-log information with seismic data, enabling the construction of a subsurface model with wide coverage and high resolution. However, training the network directly to achieve good generalization with limited well-log data as labels in a supervised manner can be very challenging. To address this challenge, we propose a pre-training and fine-tuning paradigm. In the pre-training process, we utilize a self-supervised masked autoencoder network design to derive complex feature abstractions of the seismic volumes. We then fine-tune the network by employing labeled data to construct the correlation between the seismic volumes and the well-log data. To adapt to the various types of well, we utilize 2D seismic volumes along the well path as inputs, with the well velocity value at the center of the seismic volume as the target outputs. Ultimately, we employ the fine-tuned network to estimate the velocities of the subsurface layers throughout the entire seismic area. We use two synthetic datasets and the Chevron blind dataset to test our deep learning workflow. The results indicate that the proposed workflow is effective in building a subsurface model with enhanced resolution and horizontal continuity.