Seismic Impedance Inversion Based On Deep Learning

Seismic impedance inversion is the ultimate expression form of high-resolution seismic data processing,and it is a key technology for reservoir prediction during exploration and development.However,with the deepening of exploration and development,the geological target of research has shifted from a large set of thick sand bodies to thin sand bodies.The characterization of thin sand bodies by conventional impedance inversion method not only consumes a lot of manpower and material resources,but also the accuracy of impedance obtained by inversion is difficult to meet the actual requirements.With the rapid development of artificial intelligence technology and computer hardware,deep learning method has been popularized and applied in Internet big data,and it also provides a new way to solve geophysical problems.Therefore,deep learning is used in this paper to study seismic impedance inversion.The accuracy of the linear impedance inversion method depends on the geological model,while the complete nonlinear method is expected to obtain high accuracy.In view of this,a time-domain convolutional neural network is constructed using causal expansion convolutional kernel residuals to establish the nonlinear mapping relationship between seismic data and impedance data.Then the seismic impedance samples were input for training,and the seismic data were further input into the TCN inversion mapping model to obtain the impedance.The forward data test and the application of actual data show that the TCN seismic impedance inversion method has achieved good results in the prediction of sandstone with thickness of 3-15 m.The application of deep learning seismic impedance inversion is still in the exploratory stage.It is known that while TCN seismic impedance inversion achieves good inversion effect,there are many factors affecting its inversion results.In order to accelerate the application of deep learning in seismic impedance inversion,the time-domain convolutional neural network is taken as an example to pay attention to the pre-processing of seismic data and the systematic analysis of the selection of network hyperparameters,so as to provide feasible quality control means for TCN seismic impedance inversion.At the same time,it also provides some reference for quality control means of deep learning seismic impedance inversion.The results of seismic impedance inversion based on different deep learning methods are different.In order to compare and analyze seismic impedance inversion methods under different network structures,three kinds of networks,including full convolutional neural network,convolutional cyclic neural network and time-domain convolutional neural network,are established in this paper.Firstly,network comparison is carried out to analyze its applicability.Secondly,the forward model is used to test the data,and the inversion results based on the three network frameworks are obtained.At last,well shock comparison,analysis and evaluation of impedance inversion results of actual seismic data are carried out based on logging data,in order to provide reference for the optimization of intelligent seismic impedance inversion method.It has been proved that the time-domain convolutional neural network can establish the nonlinear mapping relationship between seismic data and impedance data.However,whether the deep neural network can be used to train a inversion mapping model with good generalization ability under a small number of labeled samples remains to be explored.Based on this,this paper introduces the strategy of transfer learning on the basis of TCN seismic impedance inversion,and proposes a seismic impedance inversion method based on TCN transfer learning.This method trains the Marmousi-2model to obtain a pre-trained inverse mapping model,which can effectively predict the impedance of data with the same characteristics.However,the inversion effect is not good when it is directly used as the inversion mapping model with different data characteristics.Five Overthrust samples were added to the pre-training inversion mapping model for retraining and fine-tuning to obtain the TCN transfer learning inversion mapping model,and the Overthrust model was used as the basis for impedance inversion.The inversion result of TCN transfer learning was compared with the inversion result of 5 Overthrust sample training using TCN directly and the inversion result of Overthrust data using the above pre-training model.It was found that: The predicted impedance obtained from the seismic impedance inversion of small training samples under different data characteristics by the TCN transfer learning seismic impedance inversion method is closer to the labeled impedance,and its related evaluation indexes have also been improved to a certain degree.TCN transfer learning inversion has achieved certain results in the model test.Further,the TCN transfer learning seismic impedance inversion method is applied to the actual data.The results show that the method has certain application value in the seismic impedance inversion of small training sample data with different data characteristics.