Research On Seismic High-resolution Inversion Imaging Algorithm Based On Deep Iterative Network

Seismic full waveform inversion reconstructs the subsurface velocity model by minimizing the residuals between actual observation data and simulated data.However,the inherent nonlinearity and ill posed nature of full waveform inversion pose a significant challenge to high-resolution reconstruction.A nonlinear iterative optimization algorithm based on regularization can alleviate the ill posed nature of full waveform inversion,but there are problems such as high computational cost and the need for human intervention in setting regularization parameters.In recent years,deep learning technology has achieved the latest results in the field of inverse problems in computer vision,medical imaging,and other fields.Many scholars have made great efforts to solve seismic velocity inversion problems based on a deep neural network architecture,but most of them use data driven model based methods,and there is a problem that the inversion results depend on a large amount of data.In view of this,this paper will combine the framework of full waveform inversion iterative algorithms with deep learning to construct a set of inversion network models to alleviate the ill posed problems faced by nonlinear full waveform inversion,and can achieve high-precision imaging of velocity models.The specific research content of this article will be carried out from the following three aspects:First,this paper explores a purely data-driven depth learning inversion imaging scheme aimed at quickly reconstructing unknown velocity models from actual observed seismic records.The scheme mainly adopts the UNet++ network architecture to learn the nonlinear mapping relationship between seismic records in the data domain and velocity models in the image domain in a supervised manner.Second,this paper combines traditional non learning seismic full waveform inversion imaging algorithms with neural networks to achieve a depth learning inversion imaging scheme based on theory guidance.This scheme reconstructs an imprecise velocity model from actual observed seismic records through a traditional full waveform inversion iteration to achieve conversion from data domain to image domain,and then uses end-to-end neural networks to learn the mapping relationship between image domain and image domain,thereby achieving the goal of velocity model reconstruction.Finally,considering that the traditional full waveform inversion algorithm can be expanded into a layered network structure to solve the inverse problem,this paper proposes two model driven deep iterative expansion network methods for solving seismic inversion problems.Through automatic learning of regularization functions and regularization parameters,these two methods can avoid the limitations of algorithm applicability caused by artificial parameter selection,and can quickly and efficiently achieve high-resolution seismic inversion.Comparative experiments show that these two methods have advantages in reconstruction accuracy,convergence speed,and generalization.