Research Of Full Waveform Inversion Method In Pseudo-Depth Domain

Full Waveform Inversion(FWI)uses the residual between modeling data and observed data to establish an objective function,and uses the idea of optimization to calculate the gradient direction.Combined with the step size,the underground velocity model is updated in an iterative manner,so that the modeling data is continuously fitted to the observed data,so as to obtain accurate underground velocity model.At present,full waveform inversion has become a recognized high-precision model building tool.Although FWI has high model building accuracy,it still faces many problems in practical application.First of all,FWI is an iterative process,so the amount of calculation is huge,and it has high memory requirements.In addition,full waveform inversion method in the sense of least squares has strong non-linearity.When the initial model is poor or the data lacks low-frequency information,the error of travel time between the modeling data and the observed data is easy to exceed half cycle,resulting in the cycle-skipping phenomenon,which makes the inversion fall into the local minima.Therefore,FWI has high accuracy requirements for the initial model.Forward modeling and backward propagation of the data residual account for most of the computation of FWI.Therefore,an efficient forward modeling method can effectively improve the computational efficiency of full waveform inversion.After the velocity model is converted to pseudo-depth domain through velocity-weighted mapping,the local oversampling problem in depth-domain model is eliminated,making the longitudinal sampling more regular,which effectively reduces the number of grid points of the velocity model.Therefore,the numerical simulation method in pseudo-depth domain can effectively reduce the memory occupation and computation time of the forward modeling algorithm.In addition,the numerical simulation method in pseudo-depth domain has a low sensitivity to velocity error,and its travel time information will not change much when there is a certain degree of velocity error in the model.Based on these characteristics of the numerical simulation method in pseudo-depth domain,this paper constructed a Full waveform inversion algorithm in pseudo-depth domain(PDFWI).By introducing the pseudo-depth domain strategy into full waveform inversion,the memory and computation time of FWI algorithm can be reduced,and the accuracy requirement of the algorithm on the initial model can be reduced,so as to enhance the linearity of inversion and makes the inversion process more robust.In order to further improve the computational efficiency of full waveform inversion,this paper first analyzed the principle of CPU/GPU co-acceleration.Combined with the computation characteristics of FWI with large computation and high parallelization,the part that is suitable for parallel computation is transferred to GPU by using CUDC language.The CPU is mainly responsible for process control and data caching.Combined with CUDA programming architecture,this paper constructed the conventional and pseudo-depth domain full waveform inversion algorithms under CPU/GPU collaborative acceleration to improve the computational efficiency of the two methods.Finally,the effectiveness of the proposed method was verified by model trial,existing problems were summarized,and future research contents were prospected.