Automatic Reconstruction Method Of 3D Geological Models Based On Generative Adversarial Networks With Soft And Hard Condition Data

How to accurately reconstruct three-dimensional(3D)geological models from limited known geological survey data is one of the key challenges in the field of 3D geological modeling.Traditional methods such as deterministic modeling methods usually have a smoothing effect,which cannot accurately reflect and characterize the local details of complex geological structures.Geostatistics-based stochastic simulation methods also have limitations of the huge computational consumption and complex parameterization.Therefore,in view of the limitations of traditional methods,this thesis proposes further optimization schemes based on generative adversarial network(GAN).This work aims at the challenges of difficult adding conditional constrained data,difficult realizing the collaborative constraint of soft and hard data in the existing deep learning reconstruction methods,and difficult accurately characterizing the overall and local characteristics in the 3D geological structure.The complex heterogeneous structures are reconstructed more accurately and quickly by using the proposed methods.In a nutshell,the main innovations of this dissertation include the following points.(1)An automatic reconstruction method of 3D geological models based on the DCGAN network faced to multi-section constraint is proposed.We use multiple 2D geological sections as the conditional data to automatically reconstruct the 3D geological model.In this method,the DCGAN network model is optimized by U-Net and Patch GAN structures,and a joint loss function is proposed to guide the 3D reconstruction process.Combined with the multi-neighborhood election post-processing strategy around the attribute points to be simulated,this method can well adapt to the requirements of 3D model conditional reconstruction based on hard data of multiple sections.(2)An automatic reconstruction method of 3D geological models based on the Cycle GAN network for the collaborative constraint of section hard data and seismic soft data is proposed.In this method,the Cycle GAN network model with cyclic verification mechanism is optimized by using U-Net and Patch GAN structures,and the section loss function is added during the network training process.By using both section hard data and seismic soft data as conditional data to guide the 3D reconstruction process,this method solves the problem of weak constraint or even no constraint between sparse hard conditional data in the process of reconstruction and further improves the accuracy and rationality of the reconstruction results.Finally,a large Mesozoic terrestrial sedimentary basin in northeast China is selected as the research area of this thesis,and some areas are selected as the reconstruction experimental area.The overall 3D reconstruction results of the experimental area are completed by adopting the step-by-step reconstruction strategy according to the fixed path.The results of the application of experimental reconstruction and statistical analysis show that the proposed methods successfully reconstruct the distribution of sandstone reservoirs in the experimental area,which proves the feasibility of the proposed method for actual data in actual reservoir modeling.Comprehensive reconstruction experiments and statistical method analysis prove the effectiveness and superiority of the method proposed in this thesis.Moreover,the proposed methods have good portability and scalability for the data containing different spatial heterogeneous features,and they can be applied to other similar geological structures modeling.This work is a further improvement and development of the 3D geological modeling technology methods so that the deep learning method can effectively serve the scientific research and practical application of underground 3D structure characterization.