Details of the Abstract
| Title of paper | The Impact of Network Depth for the 2D Magnetotelluric Inversion Based on Deep Learning Inversion Models |
| List of authors | Xingong TANG, Junhu YU, Zhitao XIONG |
| Affiliation(s) | Key Laboratory of Exploration Technologies for Oil and Gas Resources of MOE, Yangtze University, Wuhan, Hubei, China, 430100 |
| Summary | This paper investigates the impact of different network depth on the performance of 2D magnetotelluric deep learning inversion models based on residual neural networks. 100,000 electric anomaly body models with different shapes and different scales generated by Gaussian random fields were design and computed. Through a self-developed batch parallel finite-difference forward modeling program based on 2D staggered grid, we conduct forward calculations and use the TE response of the models as training samples for the deep learning inversion model. Three different depths network structures, ResNet-18, ResNet-50, and ResNet-152, are trained on the samples to observe the influence of different network depths on the model performance, respectively. The results of the model inversion show that increasing the network depth within a certain range can improve the inversion performance of the deep learning model. However, beyond a certain threshold, by increasing the depth of the network, it will not lead to improvement of model performance, but increase the computational time. |
| Session Keyword | 2.0 EM theory, modelling and Inversion |
| File upload |
2.0_the_impact_of_network_dep_tang_02.pdf
|