Seismic waveform inversion best practices: regional, global and exploration test cases
Seismic waveform inversion best practices: regional, global and exploration test cases
复制标题
地震波形反演最佳实践：区域、全球和勘探测试案例
DOI：
10.1093/gji/ggw202
复制
发表时间：
2016
影响因子：
2.8
通讯作者：
Tromp, Jeroen
中科院分区：
文献类型：
--
作者：
Modrak, Ryan;Tromp, Jeroen
达到波形失拟函数的全局最小值需要仔细选择反演的非线性优化、预处理和正则化方法。由于波形反演问题容易受到与强非线性相关的不稳定收敛的影响，因此一两个测试用例不足以可靠地告知此类决策。相反，我们使用四个地震近地表问题、一个区域问题和两个全球问题来确定最佳实践。为了在方法之间进行有意义的定量比较，我们进行了数百次反转，一次改变实现的一个方面。比较非线性优化算法，我们发现有限内存BFGS在广泛的测试用例中比非线性共轭梯度方法节省了计算量。通过对预条件的比较，我们证明了一种新的由正向算子伴随矩阵导出的对角尺度比两种传统的预条件具有更好的性能。通过对正则化策略的比较，我们发现投影、卷积、Tikhonov正则化和全变分正则化在不同的情况下都是有效的。波形反演的可靠性和效率，除了策略的选择问题外，还取决于数值上的密切关注。无论选择哪种非线性优化算法，涉及行搜索和重启条件的实现细节对计算成本都有很大影响。
Reaching the global minimum of a waveform misfit function requires careful choices about the nonlinear optimization, preconditioning and regularization methods underlying an inversion. Because waveform inversion problems are susceptible to erratic convergence associated with strong nonlinearity, one or two test cases are not enough to reliably inform such decisions. We identify best practices, instead, using four seismic near-surface problems, one regional problem and two global problems. To make meaningful quantitative comparisons between methods, we carry out hundreds of inversions, varying one aspect of the implementation at a time. Comparing nonlinear optimization algorithms, we find that limited-memory BFGS provides computational savings over nonlinear conjugate gradient methods in a wide range of test cases. Comparing preconditioners, we show that a new diagonal scaling derived from the adjoint of the forward operator provides better performance than two conventional preconditioning schemes. Comparing regularization strategies, we find that projection, convolution, Tikhonov regularization and total variation regularization are effective in different contexts. Besides questions of one strategy or another, reliability and efficiency in waveform inversion depend on close numerical attention and care. Implementation details involving the line search and restart conditions have a strong effect on computational cost, regardless of the chosen nonlinear optimization algorithm.