Deep learning-based 1-D magnetotelluric inversion: performance comparison of architectures

The study compares the three deep learning approaches and assesses their relative performance solving the 1-D magnetotellurics (MT) inverse problem. MT data from a 1-D geothermal-type structure are used as an example to examine Variational Autoencoder (VAE), Residual Network (Res-Net), and U-Net architectures, adapted for 1-D MT inversion. Root Mean Square Error (RMSE) and Pearson correlation coefficient are applied as misfit measure and similarity criterion, and box plot tools are used to parameterize individual model parameters. The results show that the U-Net provides the most successful recovery of the 1-D resistivity models, even though all three approaches can produce accurate inversions of MT data. To investigate applicability of results to real data sets, the models performance are examined for the case of data containing noise. Three deep learning algorithms are robust with respect to data noise, although the U-Net is relatively superior. The study results provide a platform for more complex magnetotelluric inverse problems and ones involving real data sets.