Geostatistical Inversion Using Generative Adversarial Network (GAN)

The geophysical inversion is an important means to obtain the physical parameters of underground rocks.These lay a solid foundation for oil and gas exploration,reservoir prediction and oil and gas development.Geo-statistical inversion,as an important geophysical inversion method,has been widely applied and studied in industry and academia because of its high resolution and the horizontal continuity.The existing geostatistical inversion methods obtain prior information from well logging by constructing variation function.However,when the logging data is lacking,the prior information obtained by variation function is incomplete and inadequate.In addition,it cannot to describe the complex geologic structure,which is not effectively.To solve the above problems,this paper introduces the generative adversarial network(GAN)from deep learning,proposes a geo-statistical inversion using GAN,and realizes the algorithm of using seismic to train the network and obtain the inversion results.In this thesis,artificial neural network(ANN)is trained to extract prior information from known data,so as to solve the problem that the variation function cannot effectively extract prior information.At the same time,in order to overcome the label limit,this thesis uses the physical model to convert the generated impedance into the seismic,and then uses seismic to train the network.Finally,the geo-statistical inversion using GAN realizes the use of seismic combined with well logging well to train network,and performance better than the conventional.However,this thesis finds that the convolution kernel can only extract the features of a very small neighborhood,while the field geological features often have different relations in a large range.Therefore,this thesis introduces the self-attention(SA)mechanism into the original method,hoping to find out the potential relationship between features and enhance the pertinence of features,so as to improve the performance.The application of this method in theoretical model and field data shows that it can greatly improves the accuracy of inversion.