Inversion Of Elastic Parameters Of Tight Reservoir Based On Physically Model-driven Deep Convolutional Neural Network

Unconventional tight oil and gas reservoirs in China are widely distributed and rich oil and gas reserved,which provides the possibility for a substantial increase in oil and gas production.Therefore,the focus of oil and gas exploration and development in China has gradually shifted from conventional reservoirs to unconventional reservoirs with low porosity and low perme-ability in recent years.In addition,there are some problems such as deep burial,complex accu-mulation conditions and unclear elastic-reservoir relationship,which bring great challenges to the correct identification and prediction of oil-gas reservoirs and the characterization of fluid properties and lithology.The seismic data covers a wide range and high resolution.Estimat-ing the elastic parameters of the reservoir from the seismic data is of great significance for the identification of the sweet spots of tight oil and gas reservoirs.There are some limitations in traditional elastic parameter inversion methods,such as strong dependence on the initial model or long calculation time.Deep learning is one of the hot research directions in recent years,such methods have solved many complex problems with their powerful ability of data mining and feature extraction without relying on any physical model.Therefore,this paper takes tight sandstone reservoirs as the research object,and carries out research on the inversion of prestack seismic elastic parameters based on deep convolutional neural networks.The main work mainly focuses on:(1)According to the geological evolution characteristics and basic physical laws of the research target area,a multi-scale modeling process is proposed to generate high-quality training samples for the target area.Firstly,the logging in the work area is quality controlled and analyzed,and then the reservoir parameter volumes are obtained by adopting the geological modeling method suitable for the research target area.The cor-responding elastic parameter volumes are obtained through rock physical modeling.Finally,the seismic forward modeling method is applied to obtain the corresponding synthetic prestack seismic angle gathers.Thus,a large number of labeled seismic data volumes can be obtained,which solves the problem of few or no labels in seismic inversion.(2)In the application of field seismic data,the network seismic forward modeling are combined,the elastic parame-ters predicted by the network are further subjected to seismic forward modeling to obtain the corresponding seismic data,which is compared with the input field seismic data to form an unsupervised learning network.In this paper,the ”pre-trained +finetune” method in trans-fer learning is used to transfer the knowledge of the network trained by the synthetic training samples to initialize the unsupervised network.Due to the role of transfer learning,only a small amount of field seismic data is required to finetune the unsupervised network,and then an elastic-seismic mapping network with stronger generalization ability,better learning effect and basic geological characteristics of the target area can be obtained.(3)Evaluation of the influence of training samples on networks’ s prediction accuracy.In this paper,the effects of training samples on the accuracy of network prediction are qualitatively evaluated from three aspects: the fitness of training data and test data,the number of training samples and the diver-sity of training samples.Examples of synthetic data show that the proposed physical model-driven deep convolutional neural network has high prediction accuracy and can well describe the spatial variation of elas-tic parameters.The application results of field data show that the predicted elastic parameters by the proposed network match well with the logging,the thin interbed can be effectively identified,and the spatial variation trend of the elastic parameters can be better reflected.The examples of synthetic data and field data show that on the one hand,the network proposed in this paper has a good application effect on the elastic parameter prediction of the research target area,on the other hand,it also verifies that the training samples generated by multi-scale modeling and transfer learning have an important influence on the improvement the prediction accuracy of network.