Assess performance advantages of transformer-based super-resolution models over CNNs for fluid flows

Determine whether transformer-based deep learning architectures for super-resolution of fluid flows provide overall performance advantages compared to conventional convolutional neural network architectures, specifically in the context of reconstructing high-resolution flow fields from low-resolution inputs in fluid dynamics applications.

Background

Convolutional neural networks (CNNs) have been widely adopted for super-resolution of fluid flows and have shown success across various numerical and experimental settings. Recently, transformer-based architectures have been introduced to the same task, motivated by their strong performance in vision and sequence modeling.

Despite this growing interest, it remains uncertain whether transformers outperform CNNs for fluid-flow super-resolution. Establishing the relative performance of these architectures is important for selecting appropriate modeling tools, guiding future architecture development, and understanding trade-offs across datasets, resolutions, and flow regimes.