Robust facies classification with deep learning constrained by relative geological time

Abstract Facies classification is pivotal in the petroleum sector, facilitating reservoir characterization and enhancing strategies for hydrocarbon exploration and extraction. Recent advancements in machine learning have spurred the application of neural networks to classify facies in seismic data, marking significant progress in this field. Despite achieving commendable accuracy, these models, which process seismic waves to generate facies classifications, often encounter misclassifications at facies interfaces and within specific facies. This limitation stems primarily from the low resolution and discrete nature of seismic waveforms, compounded by restricted training datasets. Our study introduces an innovative approach by incorporating relative geologic time (RGT), which upholds the spatial continuity of seismic patterns, as an additional input to refine the encoder–decoder model. We investigate the impact of varying RGT volume accuracy on network performance, focusing on the sensitivity to errors in RGT estimation. Our findings reveal that integrating RGT data markedly enhances both the accuracy and continuity of classification outcomes. The proposed deep learning methodology, constrained by RGT, demonstrates robustness against inaccuracies in RGT estimates, offering a significant advancement in facies classification using neural networks.