Deep Wave Network for Multi-Scale Physical Dynamics Modeling

The Deep Wave Network, a newly proposed deep learning architecture, aims to model multi-scale physical dynamics in gas, fluid, and plasma systems. The research emphasizes that evaluating deep learning models solely at a single fixed size can lead to inaccurate conclusions, as different architectures display varying accuracy-cost scaling with changes in width and depth. This is especially significant for U-Net-type encoder-decoder models, known for their effectiveness in representing features across various spatial scales. U-Net achieves a multi-resolution representation through an encoder that decreases spatial resolution and a decoder that restores it for predictions, utilizing skip connections to maintain fine-scale information. While U-Net's width is often adjusted, its depth remains constant. The Deep Wave Network seeks to overcome these constraints with a new architecture that enhances the capture of multi-scale dynamics.