Research Of Time Domain Full Waveform Inversion And Its Sensitivity Analysis

To obtain accurate seismic wave propagation velocity is one of the important links in seismic data processing,and the precision of velocity parameter plays a very important role in the characterization of high precision seismic migration imaging.The conventional method of determining velocity parameters include velocity spectrum scanning and ray travel time tomography and so on.However,the conventional methods can only get a rough velocity model,which affect the accuracy of migration imaging.Full waveform inversion which can obtain the high precision velocity model is based on two-way wave equation,and the method of reconstructing the underground parameters by matching the full waveform data in the receiving record.However,full waveform inversion is a non-linear inverse problem with large computational volume and highly morbid.In particular,high precision and efficient wavefield strategy and initial velocity model with low frequency information are needed.Aiming to the problem of high precision wavefield simulation,This paper began with the accuracy and efficiency of wavefield simulation.The difference coefficients of high-order finite difference were deduced,which based on the two-way wave equation.Then,the stability condition and dispersion relation were analyzed.In order to eliminate the boundary reflection problem caused by artificial boundary,the absorption effect of Clayton absorbing boundary and perfectly matched layer(PML)absorbing boundary were analyzed.The method of wavefield simulation based on GPU/CPU parallel acceleration were studied,and the computational efficiency of full waveform inversion algorithm was greatly improved by memory and reduction algorithms.The gradient computation of full waveform inversion needs to obtain the instantaneous source wavefield at all the time and the detector adjoint wavefield accompanying the reverse transmission,and the storage is huge.In order to reduce the storage of wavefield,we studied the wavefield reconstruction strategy for storing only part of the source wavefield,which can effectively reduce the storage requirement by 90%,and alleviated the great pressure of the wave field storage of full waveform inversion algorithm.In order to alleviate the lack of lowfrequency information in seismic data,a full-waveform inversion strategy based on time second-order integration was studied.This strategy can effectively enhance long-wavelength information from seismic data.To some extent,the phenomenon of cycle-skipping caused by inaccurate initial model was avoided,and the accuracy of velocity inversion was improved.The correctness of the method was proved by using the Marmousi2 model.Finally,several numerical experiments were used to verify the influence of the initial model,the source wavelet,the observation mode and other sensitive factors on the inversion results of the whole waveform.