The Theory And Method Of Intelligent Acoustic Wave Equation Modeling And Inversion

Wave equation inversion makes full use of the kinematic and dynamic characteristics of seismic waves,which is one of the important methods to obtain the parameters of underground media.The full waveform inversion matches the simulated data with the observation data to model the parameters of the underground medium.However,due to the instability of the solution of the inversion problem and the strong nonlinearity of the inversion problem,it is still of great practical significance to improve the efficiency and accuracy of seismic wave full waveform inversion.In this paper,from the perspective of improving the accuracy of the forward and inverse results,the high-precision forward and backward method for the acoustic wave equation is studied,combined with the current popular scientific machine learning technology,the intelligent wave equation forward and backward strategy and method are proposed.Seismic modeling is an important way to obtain seismic simulation data,and the accuracy of the results directly affects the results of subsequent inversion.In the field of acoustic wave finite difference modeling,three high-precision finite difference methods were studied,namely the time-high-order finite difference method,the mixed wave number space domain finite difference method,and the center compact finite difference method.Subsequently,the study introduced scientific machine learning methods,transitioned the wave equation solution process to the training process of the neural network,and constructed the wave model form by fully constructing the neural network structure of physical perception,using the neural network automatic differential technology to introduce the wave equation residual to the loss function of the training data,the wave equation is finally solved by minimizing the loss function.The actual calculation example shows that the intelligent forward modeling method based on neural network can obviously improve the numerical simulation accuracy.In the inversion part,the full waveform inversion in time and frequency domain is first studied.Unlike traditional full waveform inversion,only the simulated wave field and the observed wave field need to be optimally matched.Reconstructed waveform inversion,this method introduces the wave equation into the objective function and reconstructs the seismic wave field enhanced by the data to meet the new objective function.It can be seen from the formula derivation that the reconstructed seismic wave field contains all the necessary information for model update,so it is no longer necessary to calculate the adjoint wave field.Model examples show that the wavefield reconstruction inversion can effectively solve the problem of inversion of the initial model lacking low-frequency information,and effectively avoid the "cycle skipping" problem.Based on this idea,research and development of seismic velocity inversion method based on physical perception neural network is different from forward neural network structure.Inversion network adds a set of velocity generation network,and introduces velocity residual information into the loss function.By optimizing the loss function,a seismic velocity model that satisfies both the wave equation and the label data.Model examples show that the global optimization algorithm based on neural network can obtain higher accuracy inversion results.However,in order to solve the computational efficiency problem of single neural network inversion,this paper finally proposes a joint inversion strategy,taking full advantage of the global optimization advantage of the neural network and the high efficiency characteristics of the traditional full waveform inversion,taking the early inversion results of the neural network as The initial model of the traditional waveform inversion,and then the subsequent inversion work is completed by the traditional inversion method.Two-dimensional complex model calculation examples show that the joint inversion strategy is not only superior to the traditional inversion method in accuracy,but also has higher computational efficiency than the single neural network inversion method.