基于神经网络的随机地震反演方法

针对随机地震反演中存在的两个主要问题,随机实现含有噪声和难以从大量随机实现中挖掘有效信息,提出了一种基于神经网络的随机地震反演方法.通过对多组随机实现及其正演地震数据的计算,构建了基于序贯高斯模拟的训练集.这也为应用神经网络求解地球物理反问题,提供了一种有效建立训练集的方法.较之传统的神经网络反演,这种训练集不仅保证了学习样本具有多样性,同时还引入了空间相关性.数值模拟结果表明,该方法只需要通过单层前馈神经网络,就可以比较有效的解决一个500个阻抗参数的反演问题.

Stochastic seismic inversion based on neural network

In view of two principal problems in stochastic seismic inversion,stochastic realizations with massive noise and being difficult to dig effective information out of mass of realizations,we propose a novel stochastic seismic inversion method based on neural network.Through the simulation of stochastic realizations and the corresponding seismic forward modeling,we build the training sets based on Sequential Gaussian Simulation (SGS).It provides an effective method to establish training sets for neural network in solving geophysical inverse problems.Compared with traditional neural network algorithms,such training sets not only have the diversity of the learning samples,but also possess the spatial correlation.The numerical simulation results show that with the aid of the proposed method,we can solve an impedance inversion problem with 500 parameters by using a feed forward neural network with only one hidden layer.