Seismic Velocity Inversion Based On Deep Learning

Oil resources are the lifeblood of national security.At present,the exploitation potential of shallow underground oil and gas resources in China is limited,and the seismic exploration targets are gradually shifting to lithologic and unconventional oil and gas reservoirs.Prestack depth migration is an important technique for oil and gas reservoir imaging in complex structural areas,and it requires high accuracy of velocity parameters.Therefore,it is of great significance to study seismic velocity inversion to improve the accuracy of velocity model for high-quality subsurface migration imaging and stratigraphic structure interpretation.Traditional velocity inversion can be divided into travel-time inversion and waveform inversion.Travel-time inversion only uses the travel-time information of seismic waves,which theoretically limits the accuracy of inversion.On this basis,waveform inversion combines the waveform information,which comprehensively reflects the kinematic and dynamic characteristics of seismic waves,and theoretically can achieve a higher accuracy.However,due to the high nonlinearity between the observed seismic data and the velocity model,the difficulty of the traditional waveform inversion method increases in the practical application.For example,the low-frequency information is often missing in the actual observed seismic data,resulting in the problem of periodic jump in the inversion process.At the same time,the reflected wave reflecting deep underground structure information is missing,which affects the accuracy of subsequent inversion.The accuracy of the initial velocity model is not enough in actual exploration,so the inversion is easy to fall into the local minimum.Therefore,it is urgent to study high precision velocity inversion and avoid or solve the above practical application problems.At present,the method based on deep learning has made rich achievements in the field of velocity inversion in seismic exploration.The effective use of deep learning method to solve many problems in the field of velocity inversion has become a popular research direction.Aiming at key problems in high-precision waveform velocity inversion,this paper uses deep neural networks and physical constraints to improve the preservation ability of seismic data features and reduce the multiple solutions of inversion problems,and studies low-frequency extension methods and velocity inversion methods of seismic data to improve inversion accuracy and efficiency.Specific research contents include:(1)In view of the problem that the lack of low frequency in the actual observed seismic data affects the accuracy of subsequent velocity inversion,this paper proposes a residual U-Net network model based on the physical attribute constraints of seismic waves to carry out low-frequency extension of seismic data.The residual block structure is used to improve the network to improve the low-frequency extension effect,and a large number of numerical experiments are designed to give a relatively good weight range of physical attribute constraints.Finally,the network can improve the recovery effect of frequency and phase while maintaining good time-domain recovery of low-frequency seismic data,and obtain high-precision low-frequency continuation seismic data,which lays a foundation for the subsequent velocity inversion.(2)To solve the initial model dependence problem,this paper proposes an attention U-Net network based on physical prior constraints for seismic velocity inversion.First,by introducing the attention mechanism and important logging data in seismic exploration combined with U-Net network,a velocity inversion subnet constrained by prior knowledge is constructed to verify the effect of the network.Then a forward RNN subnet driven by seismic wave field continuation law is constructed to realize the combination of physical law and neural network.Finally,the final velocity inversion network with physical prior constraints is constructed by connecting the inversion subnet and the forward subnet to further improve the inversion effect.