基于GPU并行的时间域全波形优化共轭梯度法快速GPR双参数反演

探地雷达（GPR）时间域全波形反演计算量巨大,内存要求高,在微机上计算难度大.本文中作者基于GPU并行加速的维度提升反演策略,采用优化的共轭梯度法,避免了Hessian矩阵的计算,在普通微机上实现了时间域全波形二维GPR双参数（介电常数和电导率）快速反演.论文首先推导了二维TM波的时域有限差分法（FDTD）的交错网格离散差分格式及波场更新策略.然后,基于Lagrange乘数法,将约束问题转化为无约束最小问题,构建了共轭梯度法反演目标函数,采用Fletcher-Reeves公式与非精确线搜索Wolfe准则,确保了梯度方向修正因子及迭代步长选取的合理性.而GPU并行计算及维度提升反演策略的应用,数倍地提升了反演速度.最后,开展了3个模型的合成数据的反演实验,分别从观测方式、梯度优化及天线频率等方面,分析了这些因素对雷达全波形反演的影响,说明双参数的反演较单一的介电常数反演,能提供更丰富的信息约束,有效提高模型重建的精度.

     

频率多尺度全波形速度反演

以二维声波方程为模型,在时间域深入研究了全波形速度反演.全波形反演要解一个非线性的最小二乘问题,是一个极小化模拟数据与已知数据之间残量的过程.针对全波形反演易陷入局部极值的困难,本文提出了基于不同尺度的频率数据的"逐级反演"策略,即先基于低频尺度的波场信息进行反演,得出一个合理的初始模型,然后再利用其他不同尺度频率的波场进行反演,并且用前一尺度的迭代反演结果作为下一尺度反演的初始模型,这样逐级进行反演.文中详细阐述和推导了理论方法及公式,包括有限差分正演模拟、速度模型修正、梯度计算和算法描述,并以Marmousi复杂构造模型为例,进行了MPI并行全波形反演数值计算,得到了较好的反演结果,验证了方法的有效性和稳健性.     

频率域全波形反演中关于复频率的研究

Laplace-Fourier域全波形反演可以利用简单的初始模型,从缺失低频信息的地震数据中得到长波长速度模型.Laplace-Fourier域全波形反演等价于本文的复频率全波形反演,但二者的实现方式不同,因此研究复频率全波形反演,可以为二者的对比研究并发展更有效的方法奠定重要基础.本文首先比较用线性增加模型作为初始模型时几个包含不同高低频成分的频率组的反演效果,再比较结合复频率之后各个频率组的反演效果,从简单模型和复杂模型的测试中都可以看出这种复频率+频率反演的方式对反演效果有明显改善.     

基于波数域梯度场分解的多尺度波形反演方法

全波形反演是一种建立高精度速度模型的有力工具,是偏移模式和层析模式的联合.然而,当初始模型较差、数据缺失低频成分和大偏移距数据缺失时,常规波形反演的层析成分更新较弱.因此,反演过程以偏移模式为主,容易导致反演快速陷入局部极小值.本文发展了基于波数域梯度场分解的多尺度波形反演方法（WGDFWI）,从梯度场中分离出层析成分,在反演的初期主要依赖层析分量更新背景速度场,为常规全波形反演建立良好的初始模型.首先,基于一种高效的隐式波场分离方法,将梯度场分解为层析成分和偏移成分.然后在层析梯度上应用二维波数域滤波器,以缓解偏移成分泄露的问题,并利用多尺度反演策略,增强反演的稳定性.利用双层模型和Marmousi模型进行试算的结果表明,该方法可以有效重构背景速度模型,为常规波形反演提供良好的初始模型,有效提高反演精度.

     

F-K域多尺度瑞雷面波全波形反演

瑞雷面波,又被称为地滚波,具有强能量和频散特性.面波对地下介质中横波速度十分敏感,利用面波测量浅层剪切波速度结构,已经在浅表勘探领域有了广泛的应用.本文研究瑞雷面波波场的F-K域全波形反演方法,成像近地表横波速度结构.首先对地震数据进行二维傅里叶变换至F-K域,以F-K域振幅谱数据构建目标函数,降低了面波波形反演的非线性;通过对目标函数的迭代求解,得到最优的反演速度模型.并且,采用从低频到高频的多尺度反演策略,减小了面波波形反演中对初始模型的依赖,缓解了周期跳跃的问题.为了验证本文方法的可行性,对水平层状模型、横向变速模型及异常体模型的横波速度结构反演,结果显示,F-K域面波波形反演方法能够有效的对近地表横波速度模型进行重构,得到高分辨率的近地表速度成像.     

基于混合自适应遗传算法的稳健全波形反演

地震全波形反演理论与技术虽已得到了广泛研究,但周波跳跃等问题的存在严重制约了常规全波形反演方法的实用化进程.基于遗传算法的全波形反演方法能够在一定程度上较好地缓解常规全波形反演面临的初始模型依赖性问题,但是当前方法仍存在收敛性和巨大的计算量问题.本文提出一种混合自适应遗传算法(HAGA),并提出基于HAGA的稳健全波形...     

频率域全波形反演方法研究进展

全波形反演是一种利用地震波传播的动力学特征来获取地下介质物性参数的反演方法,可为揭示地下精细结构提供重要依据。本文以弹性波方程作为数学模型来模拟地震波传播规律并进行相应的反演方法研究。为提高计算效率与反演结果的准确性,可将近似解析离散化（NAD）算子用于频率域弹性波方程的正演模拟。本文在频率域NAD离散的基础上推导阻抗矩阵的稀疏分块结构与反演目标函数对模型参数的梯度计算公式,由此建立基于NAD算子的频率域弹性波全波形反演方法。为验证该方法的有效性,文中通过数值实验对多种典型介质模型进行反演计算,均得到了理想的反演结果。

     

变密度声波方程多参数全波形反演策略

多参数全波形反演中各参数之间的相互耦合增加了反演的非线性程度.通过分析各参数之间的相互影响,提出合理的多参数反演策略是解决该问题的有效途径.本文从变密度声波方程出发,首先研究了密度在速度反演中的重要作用,然后分析了速度对密度反演的影响程度,进而提出了一种有利于速度、密度分步联合反演的策略.第一步,利用给定的初始模型对速度、密度进行同时反演,得到比较可靠的速度反演结果;第二步,利用第一步反演得到的速度和给定的初始密度作为初始模型,继续进行双参数同时反演,这样可以同时得到比较可靠的速度、密度反演结果.为了进一步提高反演精度,将第二步反演得到的速度、密度作为初始模型,再进行下一轮双参数联合反演.二维理论模型实验结果充分说明了本文提出的这种反演策略的有效性.     

频率域全波形反演方法研究进展

全波形反演方法利用叠前地震波场的运动学和动力学信息重建地下速度结构,具有揭示复杂地质背景下构造与岩性细节信息的潜力.根据研究需要,全波形反演既可在时间域也可在频率域实现.频率域相对于时间域反演具有计算高效、数据选择灵活等优势.近十几年来频率域全波形反演理论在波场模拟方法、反演频率选择策略、目标函数设置方式、震源子波处理方式、梯度预处理方法等方面取得了进展.目标函数存在大量局部极值的特性是影响反射地震全波形反演效果的重要内在因素之一.如果将Laplace域波形反演、频率域阻尼波场反演、频率域波形反演三种方法有机结合,可以降低反演的非线性程度.     

基于频域衰减的时域全波形反演

时域全波形反演由于采用了全频段信息,因此在迭代过程中不同波长的信息不能由低到高的逐步重建,极易陷入局部极小值.本文通过分频段的方式,对地震数据做正反傅里叶变换,利用频域指数衰减的方法逐级分离出地震数据中的高频成分,在时域上实现由低频向高频的波形反演,从而降低了反演的非线性,使不同波长的信息得到稳步恢复.同时,在高频成分衰减的过程中,后至波的能量也被削弱,由此也降低了深层反射在初始反演过程中的干扰.整个反演仅增加对数据做正反傅里叶变换过程,相较于混合域反演,无需提取全部波场的相应频率成分.在计算效率方面,利用GPU进行加速,并采用CUDA自带函数库中cufft来提高计算效率.通过对Marmousi模型测试,验证了所述方法的有效性.     

Acoustic full waveform inversion of synthetic land and marine data in the Laplace domain

Elastic waves, such as Rayleigh and mode‐converted waves, together with amplitude versus offset variations, serve as noise in full waveform inversion using the acoustic approximation. Heavy preprocessing must be applied to remove elastic effects to invert land or marine data using the acoustic inversion method in the time or frequency domains. Full waveform inversion using the elastic wave equation should be one alternative; however, multi‐parameter inversion is expensive and sensitive to the starting velocity model. We implement full acoustic waveform inversion of synthetic land and marine data in the Laplace domain with minimum preprocessing (i.e., muting) to remove elastic effects. The damping in the Laplace transform can be thought of as an automatic time windowing. Numerical examples show that Laplace‐domain acoustic inversion can yield correct smooth velocity models even with the noise originating from elastic waves. This offers the opportunity to develop an accurate smooth starting model for subsequent inversion in the frequency domain.     

基于多网格的频率域全波形反演(英文)

频率域全波形反演虽然克服了时间方向上的局部极小值问题,但是地下介质的复杂性使其在空间域仍然存在局部极小值缺陷。在优化梯度法基础上,本文采用预条件双共轭梯度稳定算法和多重网格方法计算反演中的波场传播和目标函数的梯度,在保证计算速度的同时,减小计算机内存的消耗。频率域波形反演和多重网格的多尺度性质有效改善问题极小值缺陷,加快反演的收敛速度。以局部非均匀的三孔模型和Marmousi模型的数值模拟结果验证了该算法的有效性。     

Improved hybrid iterative optimization method for seismic full waveform inversion

In full waveform inversion (FWI), Hessian information of the misfit function is of vital importance for accelerating the convergence of the inversion; however, it usually is not feasible to directly calculate the Hessian matrix and its inverse. Although the limited memory Broyden-Fletcher-Goldfarb-Shanno (L-BFGS) or Hessian-free inexact Newton (HFN) methods are able to use approximate Hessian information, the information they collect is limited. The two methods can be interlaced because they are able to provide Hessian information for each other; however, the performance of the hybrid iterative method is dependent on the effective switch between the two methods. We have designed a new scheme to realize the dynamic switch between the two methods based on the decrease ratio (DR) of the misfit function (objective function), and we propose a modified hybrid iterative optimization method. In the new scheme, we compare the DR of the two methods for a given computational cost, and choose the method with a faster DR. Using these steps, the modified method always implements the most efficient method. The results of Marmousi and over thrust model testings indicate that the convergence with our modified method is significantly faster than that in the L-BFGS method with no loss of inversion quality. Moreover, our modified outperforms the enriched method by a little speedup of the convergence. It also exhibits better efficiency than the HFN method.     

The use of low frequencies in a full‐waveform inversion and impedance inversion land seismic case study

Velocity model building and impedance inversion generally suffer from a lack of intermediate wavenumber content in seismic data. Intermediate wavenumbers may be retrieved directly from seismic data sets if enough low frequencies are recorded. Over the past years, improvements in acquisition have allowed us to obtain seismic data with a broader frequency spectrum. To illustrate the benefits of broadband acquisition, notably the recording of low frequencies, we discuss the inversion of land seismic data acquired in Inner Mongolia, China. This data set contains frequencies from 1.5–80 Hz. We show that the velocity estimate based on an acoustic full‐waveform inversion approach is superior to one obtained from reflection traveltime inversion because after full‐waveform inversion the background velocity conforms to geology. We also illustrate the added value of low frequencies in an impedance estimate.     

多主频波场的时间阻尼全波形反演

低频成分缺失和地下速度强烈变化会导致严重的周期跳现象,是地震数据全波形反演的难题.通过对地震数据加时间阻尼和时间积分降主频处理,提出了一种可有效去除周期跳现象的多主频波场时间阻尼全波形反演方法.由浅到深的速度不准确会造成波形走时失配和走时失配的累积.浅部速度的准确反演可有效地减小深部波形走时失配与周期跳现象.对地震数据施加时间阻尼得到时间阻尼数据,利用不同阻尼值的时间阻尼地震数据实现由浅到深的全波形反演.低主频波场的周期跳现象相对高主频波场的要弱.对地震波场进行不同阶的时间积分以得到不同主频的波场,把低主频波场的全波形反演结果作为高主频波场全波形反演的初始模型.应用缺失4 Hz以下频谱成分的二维盐丘模型合成数据验证所提出的全波形反演方法的正确性和有效性,数值试验结果显示多主频波场的时间阻尼全波形反演方法对缺失低频成分地震数据和地下速度强烈变化具有很好的适应性.

     

Elastic full waveform inversion with probabilistic petrophysical clustering

Full waveform inversion aims to use all information provided by seismic data to deliver high-resolution models of subsurface parameters. However, multiparameter full waveform inversion suffers from an inherent trade-off between parameters and from ill-posedness due to the highly non-linear nature of full waveform inversion. Also, the models recovered using elastic full waveform inversion are subject to local minima if the initial models are far from the optimal solution. In addition, an objective function purely based on the misfit between recorded and modelled data may honour the seismic data, but disregard the geological context. Hence, the inverted models may be geologically inconsistent, and not represent feasible lithological units. We propose that all the aforementioned difficulties can be alleviated by explicitly incorporating petrophysical information into the inversion through a penalty function based on multiple probability density functions, where each probability density function represents a different lithology with distinct properties. We treat lithological units as clusters and use unsupervised K-means clustering to separate the petrophysical information into different units of distinct lithologies that are not easily distinguishable. Through several synthetic examples, we demonstrate that the proposed framework leads full waveform inversion to elastic models that are superior to models obtained either without incorporating petrophysical information, or with a probabilistic penalty function based on a single probability density function.     

Optimized experimental design for seismic full waveform inversion: A computationally efficient method including a flexible implementation of acquisition costs

Optimized experimental design aims at reducing the cost of a seismic survey by identifying the optimal locations and amounts of sources and receivers. While the acquisition design in the context of seismic imaging applies criteria like fold, offset and spatial sampling, different attributes such as the sensitivity kernels are more relevant for seismic full waveform inversion. An ideal measure to quantify the goodness of an acquisition design relies on the eigenvalue spectrum of the approximate Hessian matrix, but this technique is computationally too expensive for practical use. A more affordable goodness measure has been proposed in the past, but we demonstrate that this measure is inappropriate for target‐oriented optimized experimental design. To address those issues, we derived a sequential receiver‐based procedure using a goodness measure based on the determinant of the approximate Hessian matrix. We show with numerical tests that it efficiently provides an optimized design for target‐oriented as well as for extensive full waveform inversion. This design allows a better reconstruction of the subsurface than an evenly spaced acquisition geometry. Furthermore, the optimization algorithm itself can easily be parallelized, therefore making it attractive for applications to large‐scale three‐dimensional surveys. In addition, our algorithm is able to incorporate variable costs, representing any kind of acquisition‐related costs, for every individual source location. The combined optimization with respect to the information content of sources and to the true cost will allow a more comprehensive and realistic survey planning and has a high potential for further applications.