Full Waveform Inversion In Time Domain And Implementation Of Optimizations

With the development of oil exploration,High-resolution imaging is expected correspondingly.Full waveform inversion(FWI)utilizes the kinematic and dynamic information of prestack seismic data to rebuild underground velocity structure.It has the potential of revealing detailed structure and lithology characteristics under complex geological background.In recent years,FWI technology has developed rapidly,and it provides a powerful support for investigating subsurface structure to enhance our knowledge of earth’s interior and to obtain detailed images of geological structures for the exploration of oil and gas deposits.FWI has become more and more popular in Geophysics,and gradually become the main means of obtaining high quality imaging.To some extent,FWI technology is an ultimate technology in the field of Geophysics,so the development of this technology is necessary in long term.FWI can be applied in time domain or frequency domain.While many researchers prefer the latter,the huge consumption of memory limits its application in field data,especially for 3D data.In term of memory,implementation of time-domain FWI is meaningful and necessary.This thesis employs time-domain FWI.and focuses on several different aspects of research,including optimization,computational efficiency,and effects of acquisition geometry on FWI and so on.The achievements of this thesis are as follows:1)Developing second order adjoint-state method in time domain.Firstly,Based on pseudo conservative form of elastodynamic equations,this thesis constructs the Lagrangian function with respect to corresponding objective function.Secondly,this thesis derives the gradient and Hessian by vector product by first and second order adjoint-state methods,meanwhile,different orders of finite different method for Hessian by vector product are given.2)Employing message passing interface technique(MPI).In order to improve the performance of modeling,this thesis uses two kinds of parallel implementations;domain decomposition and shot parallelization.The optimal combination of parallel implementations has competitive and outstanding advantage on condition of the availability of computers with enough random access memory.3)Contrast of different optimizations on FWI.The result of the deep part of inverted model is impacted deeply by optimization.The study shows that: truncated Newton and Gauss Newton perform best compared with the following optimizations,L-BFGS method performs excellently,and conjugated gradient method performs very well,but steepest descent method is always slowest.As a consequence,Newton or quasi-Newton directions play crucial role in velocity rebuilding.Truncated Newton by second order adjoint state method and finite difference have similar outstanding performance,which means that it’s sufficient to obtain Hessian by vector product by gradients of current model and gradients of adjacent model.4)Contrast of Wolfe conditions and parabolic curve fitting in line search.parabolic curve fitting performs better than Wolfe conditions in terms of the ability of step length estimation or inversion results,and adaptive step length estimation is always faster than fixed step length estimation.As a consequence,this thesis suggests that we employe adaptive parabolic curve fitting step length estimation during inversion.Based on the above research,The innovations of this thesis are as follows:1)Proposal of truncated Newton and Gauss Newton methods based on second order adjoint-state method in time-domian FWI,and implementation of truncated Newton based on finite difference in different orders.For stability,this thesis adds several tests when generating Newton direction,such as singularity test,(strong)negative curvature test and truncated test.2)Implementation of dual parallelization: combination of domain decomposition and shot parallelization.Dual parallelization greatly accelerates the computational efficiency of FWI.Domain decomposition guarantees enough random memory when storing wavefield snapshots.Shot parallelization avoids MPI communication blocking and ensures that sufficient computational nodes can be allocated.This dual parallelization structure is also suitable for implementation of 3D FWI.