Research On The Initial Model Construction Method Of Full Waveform Inversion Based On Deep Learning

In geophysical exploration,the propagation speed of seismic waves in underground media is one of the key parameters to be obtained,and obtaining an accurate initial velocity model is the most important prerequisite for achieving full waveform inversion imaging.Full waveform inversion enables accurate recovery of geological structures through full wavefield parameter information,and can even be applied to complex geological situations.However,at present,the full waveform inversion can not be well applied in the actual industry,mainly because of its strong dependence on the initial velocity model.When the initial velocity model differs significantly from the real model,the matching of the theoretical and observed wavefields is prone to period jumps.In recent years,deep learning has achieved remarkable results in many fields and can also be used in the field of seismic exploration to solve problems that are difficult to solve by traditional methods.In order to solve the problem that full waveform inversion depends on the initial velocity model,this thesis uses deep learning techniques to build the initial velocity model directly on seismic data,and investigates the frequency-domain full waveform inversion method and the full waveform inversion method based on data similarity in the real number domain.Firstly,the full waveform inversion method based on data similarity reduces the non-linearity problem in the inversion process and effectively improves the accuracy of the inversion of the deeper strata compared with the conventional frequency domain full waveform inversion results.Secondly,the U-Net network is improved by adding an attention mechanism to the U-Net network and designing the Attention U-Net network model architecture to establish the mapping relationship between seismic data and velocity model.To address the problem of low prediction accuracy of the network due to the small number of training sets,a deep migration learning method is introduced to migrate the pre-trained model to other data sets for model training.Finally,the initial model estimated by the network is used as the basis for full waveform inversion.Numerical experimental results show that the network model designed in this thesis has better performance compared with the conventional U-Net network model.The Attention UNet network model after deep migration learning has higher accuracy in modelling velocity and better reconstruction in details,proving the effectiveness and superiority of deep migration learning.The higher accuracy of the velocity model generated by this thesis’ s method compared with the inversion using a smoothed velocity model indicates that the initial velocity model built by this thesis’ s method is more accurate,and the full waveform inversion can further iterate a more accurate model,which has greater potential for industrial applications.