The Study Of MCMC Inversion Based On Convolutional Neural Network

Since entering the new century,with the increasing difficulty of resource development and the growing demand for energy in social and economic development,the industry has high requirements for new exploration technologies.Among the many technical requirements of geophysical exploration,inversion technology has always been the focus of the industry.How to develop effective inversion methods,improve the efficiency and accuracy of inversion,and accurately obtain the distribution of underground structures is great significance.At present,inversion methods can be divided into deterministic inversion and uncertain inversion,that is,probability inversion.Among them,the final inversion result of the deterministic method is a single solution,that is,the optimization result,which can not reflect the uncertain information and is easy to fall into local minima.In contrast,the result of probability inversion is a posterior probability distribution,a solution containing a large number of inversion parameters,which can obtain the uncertainty of inversion parameters and reflect the statistical information of inversion parameters.Moreover,the method of probability inversion can effectively solve the problem of local minima,get rid of the local optimal solution and find the global optimal solution.However,the probability inversion method also has its inherent problem,that is,the calculation amount is much larger than the deterministic inversion,because the probability inversion requires a large number of samples,the convergence speed is slow,and the use of full waveform data further increases the calculation amount,which makes the probability inversion of full waveform data more difficult to achieve.In order to obtain new breakthroughs in the field of geophysical probability inversion,it is necessary to introduce new technologies.In recent years,among the new technologies,the deep learning method have undoubtedly made rapid development,and have been applied to a certain extent in various fields,even made breakthroughs.The deep learning method requires a large amount of data to train the neural network,while the probability inversion method itself contains a large number of samples,which can directly provide training materials for the neural network.At the same time,the trained neural network can greatly improve the mapping speed,making more complex probability inversion possible.In this paper,the method based on Markov chain Monte Carlo(MCMC)is used to inverse GPR data,mainly through an extended Metropolis algorithm.In reality,most geophysical inversion problems are nonlinear and non-Gaussian,which makes it difficult to describe the prior probability density function.The extended Metropolis algorithm avoids solving the explicit expression of the prior probability density function,and can directly use sequence Gibbs sampling as the sample of the black box algorithm to generate prior information,reducing the difficulty of inversion and making nonlinear inversion possible.In order to solve the problem of large computation of probabilistic inversion,this paper combines probabilistic inversion based on MCMC method with convolutional neural network.In the inversion process,sequence Gibbs sampling is used to generate model samples that meet the prior requirements to avoid solving the mathematical formula of the prior probability density function.The process of solving the likelihood function of the prior sample model includes forward simulation.The trained neural network replaces the original forward mapping,which greatly improves the operation efficiency.Finally,the likelihood function of each sample model is compared through the extended Metropolis algorithm to obtain a posterior probability distribution.Because the neural network is the approximation of the forward physical formula,the forward model error is generated.In order to solve this problem,this paper uses the advantages of probability inversion to analyze and explain the model error generated by forward modeling through statistical methods.The model error is introduced into the likelihood function of inversion,so as to reduce the influence of model error on inversion results and obtain more accurate inversion results.The time-lapse inversion facing the problem of large computation and non-target areas affect the accuracy of target areas,this paper proposes a general framework for time-lapse inversion based on MCMC method,which combines the extended Metropolis algorithm with the double difference strategy.In the second inversion of the double difference strategy,the sequence Gibbs sampler realizes only the local sampling of the target area,thus reducing the amount of calculation and improving the inversion accuracy of the target area.This paper deeply studies the MCMC inversion theory,including the basic principle and algorithm implementation,and uses the depth learning network to improve the calculation efficiency of MCMC inversion.It is successfully applied to the travel time data and full waveform data of the ground penetrating radar,providing effective technical support for the realization of the practicability of probability inversion technology.The research results of this paper will play an active role in promoting fine inversion and interpretation of geophysical data.