摘要
本文使用炮并行和区域分解(物理上分割模型,使用基于MPI的分布式存储架构的计算集群,节约单个CPU内核的内存使用量,快速进行正演数值模拟)两种并行算法.该方法的每一步迭代都能确保近似海森矩阵的正定,因此,算法稳健.将时间正向传播的炮波场和反向逆时间传播的残差波场(伴随波场)进行零延迟互相关计算,得到误差泛函的梯度,然后对梯度乘以一个预条件算子,从而加快反演的收敛速度.通过抛物线搜索方法而估计步长,使用L-BFGS算法(限定内存的BFGS算法)求解模型的更新量,进行二维时间空间域弹性波全波形反演.将该反演方法应用到Marmousi2弹性波理论模型,分别反演Marmousi2理论模型的纵波速度、横波速度以及密度等三个参数.我们分别使用截止频率为2 Hz、5 Hz、10 Hz和20 Hz四个阶段的低通巴特沃斯滤波器,采用多尺度的策略,从理论模型数据的低频分量开始反演,将低频分量的反演结果作为高频分量反演时的初始模型,然后依次反演数据的高频分量.理论模型数值试验反演所得到的结果证实:二维时间空间域弹性波全波形反演计算灵活,适用于各种观测系统,能够方便地对地震数据进行加时窗;二维时间空间域弹性波全波形反演所得纵波速度模型的分辨率最高,横波速度模型的分辨率次之,密度模型的分辨率稍微差些.
This paper combines the two level parallelisms of the domain decomposition（ the domain decomposition of the physical domain will give a first level of parallelism. Using a distributed memory architecture based on MPI primitives and,on fat nodes,shared memory architecture can improve the speedup of the forward modeling） and shot parallelization double parallel algorithm on twodimensional time-space domain elastic wave full waveform inversion.Each iteration step of this method can ensure the approximate Hessian matrix is positive definite, therefore, the algorithm is robust. Zero lag cross-correlation between the forward-propagation source wavefield（ which is through finite-difference forward modelling the model） and reverse time back-propagation of the data residuals wavefield（ adjoint wavefield） to calculate the gradient of the misfit function. In order to increase the convergence speed,an appropriate preconditioning operator is applied to the gradient.Estimate the step length by the parabolic search algorithm. Update the model parameters using L-BFGS.The 2 D time-space domain elastic wave full waveform inversion method is applied to theoretical Marmousi2 model,including pressure wave velocity,shear wave velocity and density model. We use four stages to realize the inversion. Each stage has a low-pass butterworth filter,which covers different frequency ranges,with maximum frequencies of 2 Hz,5 Hz,10 Hz and 20 Hz for the numerical experiments of the theoretical model. Adopting the multi-scale strategy,which full waveform inversion start from low-frequency content to high-frequency content of the data as the inversion progresses over iterations,while keeping involved in the inversion all of the previous frequencies＇ s inversion results. Numerical experiments of the inverted results of the 2 D time-space domain elastic wave full waveform inversion calculation is flexible,which can be applied to any geometry,the seismic data can be easily added time window.The resolution of the inverted P wave velocity
作者
廖建平
刘和秀
戴世鑫
赵延林
Andrew Hursthouse
LIAO jian-ping;LIU He-xiu;DAI Shi-xin;ZHAO Yan-lin;Andrew Hursthouse(Hunan Provincial Key Laboratory of Shale Gas Resource Utilization, Hunan University of Science and Technology, Hunan Xiangtan 411201, China;State Key Laboratory of Coal Resources and Safe Mining, China University of Mining and Technology, Beijing 100083, China;Department of Geology of ttunan University of Science and Technology, Hunan Xiangtan 411201, China;School of Science and Sport, University of the West of Scotland, Paisley PAl 2BE, UK)
出处
《地球物理学进展》
CSCD
北大核心
2018年第2期671-678,共8页
Progress in Geophysics
基金
国家自然科学基金(41274126)
煤炭资源与安全开采国家重点实验室(中国矿业大学)开放基金(SKLCRSM17KFA01,SKLCRSM11KFB01)
湖南省自然科学基金(12JJ6035)
中国博士后科学基金(2012M541116)
湖南省教育厅科研项目(14C0462)
页岩气资源利用湖南省重点实验室开放基金联合资助
关键词
时间-空间域
区域分解
炮并行
二维弹性波方程
全波形反演
time-space domain
domain decomposition
shots parallelism
2D elastic wave equation
full waveform inversion
