Dilated W-Net for Geological Inversion Problems

The paper is devoted to the application of deep convolutional neural networks for geological inversion problems. The present work studies possible methods for improving the quality of neural network predictions compared to the baseline result obtained earlier using the U-net architecture with default parameters. The following techniques are considered: (a) different tensor padding methods; (b) dilated convolution to extract global features; (c) W-net architecture that follows the original ideas of U-net and extends them. A significant increase in the quality of predictions was obtained when using W-net with dilated convolution for the first sections of the network.