A Method Of The Gradient Guide For The Least-Squares Reverse Time Migration Based On VAE-LSTM

Exploration targets have become very complicated in the past few years.With the improvement of computing equipment capabilities,the current problem is how to improve imaging technology and reduce computing speed.The reverse-time migration method based on the two-pass wave equation is proposed as the most accurate migration imaging technology because it has the characteristics of not being limited by angle,it can handle all wave types and the vertical and horizontal changes of speed.But reverse time migration has a high demand for storage space,which leads to a huge I/O burden.Reverse time migration has a strong dependence on the model,and there are problems such as serious artifacts,unfidelity of amplitude,and low resolution.To achieve this goal,the least squares reverse time migration method based on the inversion theory came into being.This article deeply studied the basic theory of least squares migration and analyzed the problems existing in the conventional least squares reverse time migration algorithm.Researched and implemented the traditional least squares reverse time migration algorithm.This paper mainly discusses the gradient calculation problem of least squares reverse time migration.The specific processing method is to optimize the gradient calculation process with the Recurrent Neural Network.A method based on VAE-LSTM is proposed to construct the gradient of least squares reverse time migration.In order to solve the above problems,this article mainly did the following work:1)Research the least squares reverse time migration method systematically,and introduce the theoretical basis of the least squares reverse time migration method and the iterative solution method of the least squares reverse time migration;2)Perform numerical trial calculations on the conventional least squares reverse time migration method,analyze the problems existing in the conventional least squares reverse time migration method and propose methods to solve the problems;3)Introduced the Recurrent Neural Network(RNN).In order to solve some of the shortcomings of the RNN,the Long and Short-term Memory Neural Network(LSTM)was introduced,at the same time,we introduced the Variational Autoencoder(VAE).Finally,the common loss functions in the training process are summarized;4)By building a long and short-term memory neural network framework,then use it as the encoder part of the variational autoencoder to project the LSRTM model into the latent space of the VAE for calculation,and finally achieve the purpose of self-learning and generating migration results,so that the calculation efficiency of the least squares reverse time migration method can be improved.