Deep learning-based seismic lithofacies prediction in sparse well areas via geology-informed pseudo-well construction and transfer learning

Abstract

Accurate lithology prediction from seismic data plays a critical role in unraveling the complexities of subsurface geology, enabling informed decision-making in geo-energy exploration and production, geological storage of CO2, and geological hazard assessment. Deep learning approaches, with their capabilities of feature extraction, mapping non-linear relationships, and handling high dimensional features, show great potential in seismic reservoir characterization. However, limited well logging data due to high drilling costs pose challenges for deep learning model training, particularly in frontier exploration and early development stage. Existing data augmentation methods often focus on increasing data quantity without effectively utilizing geological knowledge, potentially limiting their ability to capture the realistic complexity of data. To address this challenge, especially in sparse well regions, we propose a geostatistics-based pseudo-well construction methodology. By considering geologic stratification, the lithofacies are simulated using the Markov chain method, and the corresponding elastic features are simulated using sequential Gaussian simulation. This methodology enhances the reliability and accuracy of pseudo-well construction, with more geological consistency with the actual wells. Then, using the limited actual well data, we use transfer learning strategy to predict lithofacies from prestack data and seismic inversion via supervised convolutional neural network. We employ the proposed methodology in a coal-bearing clastic reservoir. Based on the blind well test, the strategy of combining pseudo-well data and transfer learning leads to a notable enhancement in the F1 score of sandstone from 57.45 % to 62.16 %, as well as an overall F1 score improvement from 52.92 % to 57.89 %. We apply this method to 2D seismic profiles (prestack data and inversion results), and the predicted spatial distribution of the lithofacies shows better agreement with the lithofacies in actual wells and more geological reasonableness.