Fault-karst reservoir is of a special type distributed in the Ordovician strata in the Tarim Basin, China. It’s characterized by deep burial, complex genesis and strong heterogeneity. Due to sparse well data and low seismic quality and other adverse conditions, its accurate characterization and fine modeling are faced with great challenges. In the study, an integration of drilling, core, outcrop and 3D seismic data is applied to build a deep learning-based training dataset for the fault-karst reservoir with the guidance of architecture mode of fault-controlled fractured-vuggy reservoir. Based on the comprehensive analysis of deep learning network, we propose a deep learning-based modeling method suitable for fault-karst reservoirs. The results show that the “in-situ, equal-scale” training dataset established based on multi-source data is the basis for deep learning-based modeling of fault-karst reservoirs. The selected pix 2 pix (P2P) neural network could realize the 3D model prediction of fault-karst reservoirs by seismic data. A 3D faultkarst reservoir model is then established for the south segment of the No. 5 fault zone in Shunbei area following the built of training network. The model is conformed to the geological mode and distribution pattern of the reservoir type on all fronts, and also highly consistent with the reservoir prediction based on drilling data. One of the key research directions therefore lies in improving the accuracy and conditional degree of deep learning-based geological modeling of fault-karst reservoirs.

Problems including insufficient quantity and low resolution of data face exploration and development of deep and complex reservoir targets, and the traditional geological modeling methods have been inadequate in terms of technical needs. The intelligent geological modeling represented by deep learning capable of fully integrating multiscale and multi-dimensional data as well as expert knowledge, is a key research and development direction of geological modeling technology. The study discusses the deep learning-based geological modeling driven by sedimentary process simulation following the comprehensive analysis of the advantages and disadvantages of stratigraphic forward modeling and deep learning-based geological modeling technology. First, forward modeling of sedimentation is carried out based on comprehensive geological analysis, parameter uncertainty is analyzed, and a large amount of geological models are established through parameter disturbance as a training dataset; Second, the geological patterns contained in the learning dataset are learned with the conditional Generative Adversarial Nets (cGAN), in which the Generative Adversarial Networks (GAN) takes the conditional data such as well and seismic data as the input, and the geological model as the output; Finally, the trained GAN is applied to the real conditional data to obtain the geological model of the target block. The feasibility of this method is verified through testing on the typical geological profiles of the main block of Puguang gas reservoir, and the impact of the training dataset scale on simulation results is analyzed. The combination of sedimentary simulation and deep learning could make up for the shortage of training data and indirectly realize knowledge-driven deep learning-based geological modeling. The method is therefore of great significance to popularization.