Details of the abstract
|Title of paper
|2D U-Net convolutional networks for 1D inversion of magnetotelluric data
|List of authors
|Mehdi Rahmani Jevinani, Banafsheh Habibian Dehkordi, Mohammad Hossein Rohban and Ian J. Ferguson
Institute of Geophysics, University of Tehran, Tehran, Iran, Institute of Geophysics, University of Tehran, Tehran, Iran, Department of Computer Engineering, Sharif University of Technology, Tehran, Iran, Dept. of Geological Sciences, University of Manitoba, Winnipeg, Manitoba, Canada, Ij.Ferguson@umanitoba.ca
Deep convolutional networks as one of the main developments in deep learning approaches have been applied increasingly in geophysical inverse problems in recent years. The capabilities of 2D U-Net convolutional network for 1D inversion of magnetotelluric data are examined in this study. We apply an innovative method to generate the large number of sample data that are generally required by deep learning-based approaches. Layers with fixed and variable thicknesses and (apparent resistivity) data with and without phases are considered. From the total of 1 million data samples, 95% of them are used as training and validation sets and 5% as test data. Input data are scaled for training and rescaled to form output data or predictions.
The original 2D U-Net is used for segmentation problems. To 1D inverse modeling of magnetotelluric data, a 2D U-Net with 36 layers is designed including two main modifications: First, at the beginning and end of the compression and decompression path, filter 32 is replaced by filter 64. Second, instead of tow, four convolutional layers are used for each filter. The kernel size is 3×2 and batch normalization and ReLU activation functions are considered. Dropout ratios of 0.5, 0.3 and 0.1 are also added to prevent overfitting. Finally, the Adam algorithm with learning rate of 0.0001 and Huber loss are used as optimization algorithm and loss function, respectively.
In the random design of one-dimensional models, there is a possibility of creating layers that due to the values of their electrical conductance cannot be easily identified in magnetotellurics. Therefore, subsurface resistivity values as outputs of the deep learning model are also recovered with similar limitations in resolution.
|2.0 Theory, Modelling and Inversion