基于最优输运与循环神经网络的全波形反演  

Full waveform inversion based on optimal transport and recurrent neural networks

作　　者：李燕梅 顾焕申 许凯[2] 孙振涛[2] 王文龙[1] LI YanMei;GU HuanShen;XU Kai;SUN ZhenTao;WANG WenLong(College of Mathematics,Harbin Institute of Technology,Harbin 150001,China;China Petroleum and Chemical Corporation,Sinopec Geophysical Research Institute,Nanjing 211103,China)

机构地区：[1]哈尔滨工业大学数学学院,哈尔滨150001 [2]中国石油化工股份有限公司石油物探技术研究院,南京211103

出　　处：《地球物理学进展》2022年第6期2408-2416,共9页Progress in Geophysics

基　　金：国家自然科学基金(41804108)资助。

摘　　要：针对全波形反演收敛慢与对低频速度模型的依赖等问题,本文将深度学习与最优输运理论结合.利用循环神经网络求解时间域波动方程,进而实现波动方程的正演模拟与波形反演.在目标函数的建立上,以最优输运为基本理论,比较了不同变换下的目标函数对反演效果的影响;在收敛效率上,利用深度学习中的小批量优化算法,对比不同批量数据下的反演效果.数值实验表明,线性正变换优于比较范围内的其他变换,可有效提高反演精度.对比常规全批量的全波形反演,小批量的随机反演策略在同等计算量情况下明显提高了收敛效率.Aiming at the problems of slow convergence and high dependence on low-frequency velocity models in full waveform inversion,we combine deep learning and optimal transport theory.Recurrent neural network is used to solve the wave equation in the time domain.Thus,the forward simulation and waveform inversion are realized.In the establishment of the objective function,the optimal transport is used as the basic theory,we compare the influence of different transformations in the objective function on the inversion;in terms of convergence efficiency,the mini-batch optimization algorithm in deep learning is used to compare different batch sizes.Numerical experiments show that the linear forward transformation is better than other transformations in the comparison range,which can effectively reduce the dependence on the initial model.Compared with the conventional full-batch full waveform inversion,the small-batch stochastic inversion strategy improves the convergence efficiency with the same amount of computation.

关 键 词：全波形反演 最优输运 循环神经网络 

分 类 号：P631[天文地球—地质矿产勘探]