Research Of Model-Data-Driven AVO Inversion Method

Velocity and density are key parameters in seismic exploration and play important roles in seismic data processing and interpretation.Improving the accuracy and resolution of velocity and density is conducive to the accurate and fine identification of reservoir lithology and fluid.Model-driven AVO inversion and data-driven seismic parameter prediction are important methods to obtain velocity and density from seismic data.The former usually starts from the initial parameter models,uses the error between synthetic seismic data and real seismic data to construct the objective functions,and iteratively updates the parameter models with the least square optimization algorithm or the Bayesian algorithm until the objective functions tend to converge to realize the stable inversion for velocity and density.Limited by the bandwidth of real seismic data,the methods usually need proper initial low frequency models,and cannot provide high-resolution inversion parameters.The latter usually starts from the initial neural networks,constructs the objective functions by using the error between the prediction parameters and the sample parameters,iteratively updates the neural network weight coefficients with the back-propagation algorithm until the objective functions tend to converge,and realizes the prediction for velocity and density based on the trained neural networks.Compared with the model-driven AVO inversion method,which mainly uses the seismic data with limited bandwidth,the data-driven seismic parameter prediction method mainly uses the well-log data with larger bandwidth and can provide higher resolution velocity and density.However,this kind of method usually needs many representative training samples,which is suitable for large sample data.There are only a few wells in the work area,which cannot provide enough well-log data and training samples and are seen as small sample data.It limits the application of the data-driven seismic parameter prediction method.To solve to above problems,this dissertation mainly focuses on the high-precision and high-resolution inversion methods of velocity and density.The main achievements are as follows:Firstly,a model-driven high-order AVO inversion method based on the series reversion is proposed in this dissertation.The method linearizes the nonlinear inversion process of the high-order approximate formula by using the series reversion,which can solve the low-accuracy problem of the conventional linear AVO inversion method with high efficiency.However,limited by the bandwidth of seismic data,the resolution of inverted velocity and density is not high.Secondly,a "three-step" model-data-driven AVO inversion method based on the GRU neural network is proposed in this dissertation.To break through the limitation of seismic data bandwidth and improve the inversion accuracy and resolution of parameters,in the method,the GRU neural network is used to construct the relationship between the seismic data and the seismic parameters.The F-norm is used to calculate the error between the prediction parameters and the sample parameters,the error between the low-frequency component of the prediction parameters and the initial low-frequency models,and the error between the synthetic seismic data and the real seismic data.Based on these three errors,neural network training,neural network optimization,and neural network inversion are realized in turn.Tests on model data and field data show that,compared with the model-driven AVO inversion and the data-driven seismic parameter prediction,the "three-step method" can provide inversion parameters with higher accuracy and resolution.Thirdly,a "two-step" model-data-driven AVO inversion method based on the U-net neural network is proposed.Because the “three-step method” updates the weight coefficients of the neural networks three times,it needs a lot of calculation time and memory.In addition,it uses the F-norm to calculate the error between the synthetic seismic data and the real seismic data.When the amplitude difference between them is large,the inversion accuracy of the method is not high.Furthermore,although the GRU neural network is suitable for processing time-series data,it has high requirements for the quality and quantity of training samples.So,it is difficult to effectively process small sample data,which affects the accuracy of seismic parameters inversion to a certain extent.To further improve the inversion efficiency and accuracy of velocity and density,the “two-step method” uses the U-net neural network to construct the relationship between the seismic data and the seismic parameters;applies the F-norm to calculate the error between the prediction parameters and the sample parameters,and the error between the prediction parameters’ low-frequency component and the initial low-frequency models;utilizes the zero-lag cross-correlation function to calculate the error between the synthetic seismic data and the real seismic data.These three errors are coupled into two objective functions,which are used to realize neural network training and neural network inversion in turn.Tests on model data and field data show that compared with the "three-step method",the "two-step method" can obtain higher precision velocity and density in a shorter time.Fourthly,a "one-step" space-variant objective function-based model-data-driven AVO inversion method is proposed.The “three-step method” and the “two-step method” have a common problem,that is,the constraint weights of training samples on all inversion target CDPs are the same,which is not in line with the fact and affects the accuracy of inversion.Because the training samples are mainly generated by the well-log data.According to the similarity theory,the closer the data to the well is,the greater of similarity between the well-log data and inversion parameters is,the greater the constraint weight of well-log data on its inversion is,and vice versa.Therefore,the constraint weights of training samples on inversion should change with the distance from the wells.In addition,although the "two-step method" is one step less than the "three-step method",it still needs large computing memory and time.Moreover,T U-nets used in the "two-step method" contain up-sampling and down-sampling structures,which will lose some detailed information about structure and reservoir in the process of parameter inversion and affect the accuracy of inverted parameters.To solve the three problems,the “one-step method” improves the U-nets in the "two-step method".The three errors are calculated by the F-norm and the zero-lag cross-correlation function,respectively,and then are coupled into a space-variant objective function for neural network inversion.The construction of the space-variant objective function follows the idea of inverse distance weighting.Tests on model data and field data show that the inversion accuracy and efficiency of the "one-step method" are higher than those of the "two-step method".Fifthly,a model-data-driven AVO inversion method based on the invertible neural network is proposed.The above four methods have a common feature,that is,they need proper initial low-frequency models.However,it is difficult to provide such initial low-frequency models in some work areas,which limits the application of these methods.To solve the problem that the inversion method is highly dependent on the initial low-frequency models,the new method uses the invertible neural network to learn the definite forward mapping from the seismic parameters to the seismic data and then relies on its bijective structure to realize the stable inversion of velocity and density.The method mainly uses the parameter data randomly generated according to the data characteristics of the work area,rather than the proper initial low-frequency models or accurate training samples.Tests on model data and field data show that the method can get high-precision low-frequency and medium-frequency velocity and density without the initial low-frequency model and has good practicability and advanced nature.