Patch-Based 3D Variational Autoencoder for Super-Resolution of Turbulent Channel Flow
A recent paper on arXiv has introduced an innovative machine learning technique aimed at enhancing the super-resolution of turbulent flows. This approach employs a patch-based three-dimensional variational autoencoder (3D-VAE) to derive high-resolution flow fields from lower-quality data, tackling the challenges posed by direct numerical simulation (DNS) at elevated Reynolds numbers. The model generates a local high-resolution block of 16^3 from a broader coarse neighborhood and utilizes convolutional application across the entire domain, ensuring that the parameter count remains unaffected by domain size. By training on streamwise velocity data from a single DNS snapshot of turbulent channel flow, this method seeks to address the shortcomings of current 2D techniques that do not effectively capture vortex stretching. The paper can be found on arXiv under identifier 2507.22082.
Key facts
- The paper proposes a patch-based 3D variational autoencoder for super-resolution of turbulent channel flow.
- The method reconstructs a local 16^3 high-resolution block from a larger coarse neighborhood.
- The learned operator is applied convolutionally across the domain with overlap averaging.
- Parameter count depends only on patch size, not domain size.
- The model is trained using streamwise velocity from a single DNS snapshot of turbulent channel flow.
- Direct numerical simulation (DNS) becomes prohibitively expensive as Reynolds number increases.
- Existing super-resolution methods focus on 2D data and extend poorly to 3D.
- The paper is published on arXiv with identifier 2507.22082.
Entities
Institutions
- arXiv