Research On High-performance Computing Methods For Time-domain FWI

Full waveform inversion(FWI)technique has the capacity to portray structural details and directly reconstruct multiple seismic lithology parameters using prestack data.The study of FWI is a complicated system which concerns the forward modeling,gradient estimates,parallel task partition strategies and so on.For a long time,the low computing efficiency has been a persistent headwind to the FWI research and industrial applications.Time domain full waveform inversion is one of the most important research directions in seismic nonlinear inversion.This paper focuses on how to improve the efficiency of waveform inversion from three points of view.First,from the aspect of improving the forward algorithm,the dissertation separately introduced the time/space dimensions differential discretization of the common first-order and second-order wave equations,as well as the corresponding PML and NPML boundary conditions.We reach a conclusion that traditional finite difference method is computationally redundant,hence this paper propose to use the information of the first arrival travel time to determine the effective range of the finite difference method,namely first-arrival time domain finite difference method(FATD-FDM)forward modeling.FATD-FDM can reduce the computing complexity about 30%.Second.The second aspect is inversion algorithm.This paper analyzes the reasons causing uneven gradient energy distribution and artificial noise,and take steps to fix gradient field by illuminating,filtering and regularizing to reduce the inversion failure probability.In addition,this paper introduced the positive effects of multi-scale strategy and laplace wave field attenuation on reducing the inversion nonlinearity and avoiding the objective function trapped in local extremum.As for the last aspect parallel computation technology,this paper made a detailed discussion on the basic principle of high performance computing in all from hardware to software and divided FWI into three levels,which are parallel between/among spatial grids,parallel between/among shots as well as parallel between/among single-shot gradients.Therefore,the MPI+OpenMP+SIMD/SIMT three levels heterogeneous parallel scheme is projected in order to accommodate characteristics of FWI.As for the problem of large data storage for full wave field,the existing strategies are solvable but still have some defects.For example,disk dynamic storage has disk access delay,random boundary storage will increase the amount of forward computation and change space in time.This paper points out that gradient imaging requires only the first wave information of the positive wave field rather than the full wave field information.Based on this principle,this paper use the first arrival travel time information to compresses the full wave data almost 2 to 3 orders of magnitude on the premise of correct gradient image.In short,a set of economical and efficient FWI scheme is initially formed through multimethod acceleration and wave field data compression,which has high parallel efficiency,high speedup ratio and outstanding extendibility.