Construction of a multi-source fault-controlled karst cave training dataset and application in three-dimensional modelling using generative adversarial networks

Currently, there is no comprehensive training dataset available for the modelling of fault-controlled karst caves using deep learning. In this study, we constructed prototype models for fault-controlled karst caves using outcrop data, seismic data, reliable geological models, and object-based methods. We combined, rotated, cropped, and selected prototype models from different sources to create a reliable and diverse training dataset for fault-controlled karst caves. Additionally, we constructed corresponding virtual well and probability map training datasets, all of which were used to train conditional generative adversarial networks (GANs). The trained generator convolutional neural network was applied to TH12330 well block, Tahe Oilfield. The generated multiple geological models for fault-controlled karst caves were consistent with geological patterns, conditioning well data, conditioning probability map data, and aligned with fracture structures, fractures, and cumulative production. This research explores the construction of a multisource training dataset for fault-controlled karst caves and has achieved significant success in a real application example. Furthermore, it provides new insights into building reliable and diverse training dataset for deep learning modelling in other types of reservoirs.