基于深度降噪自编码神经网络的中国大陆地壳厚度反演

Inversion of crustal thickness in Chinese mainland based on deep denoising autoencoder neural network

  • 摘要: 本文采用基于数据驱动的深度降噪自编码网络构建了瑞雷面波群速度、相速度频散特性与地壳厚度的正反演函数关系,并利用最新频散模型反演了中国大陆的地壳厚度。对于神经网络架构体系的评价,除了考虑传统意义上的测试误差、训练误差之外,本文还用已知物理原理的正演结果与网络预测结果进行比较;在设计网络构架时,同时考虑地球模型和面波频散的正反演问题,即解码过程对应正演过程,编码过程对应反演过程。另外,针对观测频散数据包含噪声的特点,对训练样本加噪声,使解码器解码出无噪声输入,以达到对观测数据降噪的目的。对网络各种参数多次调试、分析再优化组合,最终获得稳健的神经网络,并据此反演出中国大陆的地壳厚度。本研究结果与已有的不同手段得到的地壳厚度模型的吻合度较高,表明深度降噪自编码神经网络能很好地揭示面波频散与地壳厚度之间的非线性关系,是利用面波频散反演地壳厚度的一种可行的和可信的方法。

     

    Abstract: Chinese mainland is a composite continent composed of many micro-blocks, fold belts and orogenic belts after evolution over a long geological period. Chinese mainland crust is of complex crust-mantle structure, and crustal thickness is one of the most important parameters. At the same time, Rayleigh surface wave group velocity and phase velocity have strong constraints on the crust and upper mantle structure. In this paper, a data driven, named deep denoising autoencoder (DDAE) neural network is used to explore the relationship of forward and inversion between Rayleigh wave group velocity and phase velocity and crustal thickness, and we invert the crustal thickness of Chinese mainland by the latest dispersion model. In this paper, the evaluation of neural network architecture is put forward. In addition to the traditional test error and training error, it is compared with the result of network prediction with the forward process of the known physical principles. When designing the network architecture, the forward and inverse problems of the earth model and the surface wave dispersion are considered simultaneously, that is, the corresponding forward process is decoded, and the encoding process corresponds to the inversion process. At the same time, in view of the noise characteristics of the observed dispersion data, the training sample is polluted by noise, so the decoder decodes the noise-free input to denoising the observed data. After debugging, analyzing and optimizing the various parameters of the network, a robust neural network was finally obtained, and the crust thickness of the Chinese mainland was reproduced accordingly. The results of this study are in good agreement with the crustal thickness models obtained by different methods, suggesting that the deep denoising autoencoder neural network can well reveal the nonlinear relationship between surface wave dispersion and crustal thickness, and it is a feasible and credible method to solve the problem of surface wave dispersion inversion for crustal thickness.

     

/

返回文章
返回