Research On Velocity Modeling Method With Multiple Information Constraints

Oil and gas are important energy minerals and strategic resources,playing an extremely important role in the national economy.Seismic exploration is currently the most powerful ways for the discovery of oil and gas resources.It processes and analyzes the differences in the propagation state of seismic waves in different media to obtain underground structural and lithological information.In seismic exploration,obtaining seismic wave velocities that directly reflect physical parameters is the core technical.Its role runs through the entire seismic data processing process and is a prerequisite for establishing high-precision seismic imaging,reservoir description,geological interpretation,and other technologies.Oil and gas exploration work is facing increasingly complex structural environments,which require higher accuracy in velocity modeling.In addition to seismic data,making full use of diversified information such as geology and logging,comprehensively considering the rich structural information contained in migration imaging and seismic interpretation,and conducting targeted constraints on the velocity modeling process are powerful strategies and important research directions for reducing inversion multiplicity,improving inversion accuracy,and obtaining highprecision velocity models that are more consistent with real geological structures.The near-surface velocity structure is more complex compared to the deep part,often manifested as obvious elevation fluctuations and rapid lateral changes in the medium,which have a significant impact on the propagation of seismic waves.Establishing an accurate near-surface velocity model is the foundation for subsequent seismic data processing.First arrival traveltime tomography method has the advantage of stable inversion,and is widely used in near-surface velocity modeling.However,due to the limited information contained in the first-arrival traveltime,the inversion results are prone to multiple solutions.At the same time,the computational process requires mesh generation of the model,which indirectly reduces the accuracy inversion and limits the fusion ability with multi-scale information.In this paper,we embed the Eikonal equation into a physics-informed neural network instead of the forward process,and use an automatic differential algorithm instead of the inversion process to achieve automated tomography.Further,microlog data containing rich near-surface information is introduced into the loss function to effectively fuse seismic and microlog data to constrain the inversion process,forming a multi-source information fusion tomography method based on physics-informed neural networks.Numerical tests show that our method can improve the spatial resolution of velocity models without providing accurate initial models,and obtain near-surface velocity models that are more consistent with the actual geological structure.Stacking velocity is used for the initial imaging of seismic data in the time domain,which is an important means to obtain the overall underground velocity and lays a solid foundation for further depth domain imaging.Picking velocity semblance is the core step of stacking velocity analysis.In field seismic data,random noises and multiples can cause strong energy cluster interference in the velocity semblance,which cannot be accurately distinguished from the energy cluster of primary waves.Moreover,traditional manual picking methods have low efficiency in large-scale data processing and are susceptible to interference from multiple factors.In this paper,we propose an automatic velocity analysis method with physics-constrained optimal surface picking,which arranges the 2D velocity semblance into a 3D velocity semblance volume by common middle point(CMP)location.Based on the dynamic programming algorithm,physical constraints on layer velocity are added in the time direction to attenuate abnormal energy clusters,and slope smoothing constraints are introduced in the CMP direction to enhance the spatial structure consistency of the velocity field,so as to achieve automatic and efficient picking.Numerical tests show that our method can highlight the energy of primary waves and obtain more accurate stacking velocity that consistent with true geological knowledge.Field data test shows that our method can obtain a stacking velocity field with structural consistency,thereby improving the quality of the stacking image.The depth domain velocity model directly corresponds to the real underground structure,making up for the limitations of time domain velocity analysis in complex geological environments and severe changes in lateral media,directly determining the accuracy of seismic data interpretation.Deep domain velocity modeling includes two main steps:establishing the initial model and updating the residual velocity.Whether the common image gather(CIG)are flattened is an important criterion for determining velocity accuracy.In this paper,we investigate the improvement of depth domain velocity modeling accuracy from three aspects:establishment of initial model,determination of reflection points,and picking of imaging gathers.We first apply optimal surface picking to the 3D residual curvature semblance volume based on residual curvature analysis method,and obtain an initial model with lateral construction consistency by adding smooth slope constraints in depth and common image point(CIP)direction.Further,we extract the maximum value of the picking path corresponding to the residual curvature semblance as the position of the reflection point.The CIG can then be globally optimal picked based on the dynamic programming algorithm,using previously selected positions as seed points,which results in automatic high-precision picking of reflection events.Field data test shows that our method can obtain relatively stable results,providing reliable data picking support for tomographic inversion.Full waveform inversion(FWI)is the theoretically most accurate method for deep domain velocity modeling.It makes full use of the effective dynamic and kinematics information in seismic data,and gradually reduces the difference between actual observation and theoretical forward modeling data through iterative inversion of optimization methods,becoming a key research direction in geophysics.FWI has extremely strong nonlinearity and is prone to cycle skipping due to the influence of initial model and data quality,resulting in inversion results that fit the data but do not have practical geological meaning.In this paper,we first perform pre-stack depth migration based on the low wavenumber initial velocity model,and construct an implicit structural model using migration images and interpretation information.Further,global constraints are applied to FWI based on the implicit structural model,and the inversion solution space is compressed by optimizing gradient parameters to obtain a velocity model with practical geological meaning.Numerical tests show that our method is not easily trapped in local minima and can still obtain stable and reasonable inversion results even when the initial model is inaccurate.Field data test shows that our method has high inversion resolution in complex geological areas and high consistency with the structural information in migration images.