Geostatistical Inversion Method Based On Deep Learning

In the stage of oil and gas field development,how to obtain precise and accurate geological model is the key and difficult point of reservoir prediction.Traditional two-point geostatistics method characterizes reservoir heterogeneity based on Variogram,but it is difficult to reproduce reservoir structure with complex sedimentary facies association in space.Multipoint geostatistics,which takes training image as the core and represents spatial correlation of multiple points,the spatial geometry and distribution pattern of geological bodies can be characterized.The training image is a geological model that can quantify the geological model of the study area.By extracting quantitative prior geological information from the training image,it can be applied to inversion modeling under the constraints of conditional points.However,the large amount of data events in the training image results in low computational efficiency,and its non-linear structural characteristics are difficult to reflect.Aiming at the disadvantage of multi-point geostatistical inversion,in this paper,the trained image features are learned by Generative Adversarial Network based on the deep learning algorithm,and the trained neural network is used for inversion modeling.On the basis of completing the algorithm,the SGAN method is tested on several training images,and the quality of the trained neural network is evaluated by analyzing the results of SGAN generation.The results show that the model generated by this method has high similarity with the trained image,and the generation of thousands of models only takes several seconds.The efficiency of the model generation is more breakthrough than that of the conventional method.The trained neural network is combined with MCMC-Bayes inversion method,is applied to the geological modeling under the delta sedimentary background based on the conditional data of the study area.The inversion results can reflect the combination pattern of training images on the basis of matching the conditional data,and obtain a posterior probability volume model through multi-chain parallel operation in a very short time.This method will play a positive role in geological Inversion Modeling and has great influence on reservoir description.It will provide high quality geological model for exploration and development of oil and gas more efficiently.