1-D Multi-Parameter Inversion Based on Deep Neural Network for Geophysical
- Resource Type
- Conference
- Authors
- Liang, Bingyang; Yang, Shengpeng; Zhou, Yuanguo; Yu, Simiao; Gong, Yubin
- Source
- 2023 Cross Strait Radio Science and Wireless Technology Conference (CSRSWTC) Cross Strait Radio Science and Wireless Technology Conference (CSRSWTC), 2023. :1-3 Nov, 2023
- Subject
- Computing and Processing
Fields, Waves and Electromagnetics
Photonics and Electrooptics
Signal Processing and Analysis
Wireless communication
Training data
Artificial neural networks
Conductivity
Nonhomogeneous media
Numerical models
Green's function methods
Multi-parameter inversion
DNN networks
Dyadic Green's function
- Language
- ISSN
- 2377-8512
In this paper, a new inversion method is proposed to reconstruction of conductivity and layer thickness of layered media based on machine learning deep neural networks. A full connection deep neural network with 10 hidden layers is used. Training data are generated by Dyadic Green's function of layered media. The hidden layer of the network uses LeakyReLU function as the activation function. Numerical examples of true value of model are provided to benchmark the performance of the proposed methods. The results show that the DNN method can reconstruct the conductivity and layer thickness parameters of the layered model, and the reconstruction results show that the BP neural network has better convergence and accuracy. For the trained network, multi parameter model results can be reconstructed within 0.01s.