Patch-Based 3D Variational Autoencoder for Super-Resolution of Turbulent Channel Flow

Abstract:Direct numerical simulation (DNS) accurately resolves all spatio-temporal scales of wall-bounded turbulence but becomes prohibitively expensive as the Reynolds number increases. Super-resolution (SR) provides a practical alternative by reconstructing fine-scale flow structures from coarse fields. Most existing SR methods focus on two-dimensional data, where vortex stretching is absent, and extend poorly to three dimensions because model complexity increases with the reconstructed volume. We propose a patch-based three-dimensional variational autoencoder (3D-VAE) that reconstructs a local (16^3) high-resolution block from a larger coarse neighbourhood. The learned operator is then applied convolutionally across the domain with overlap averaging, making the parameter count dependent only on patch size rather than domain size. The model is trained using the streamwise velocity from a single DNS snapshot of turbulent channel flow at (Re_\tau \approx 1000) from the Johns Hopkins Turbulence Database and evaluated on a held-out snapshot. Compared with DNS, the proposed method achieves a mean absolute error of 0.055, outperforming tricubic (0.075) and Lanczos (0.076) interpolation. In spectral space, it reduces the mean absolute error of the two-dimensional Fourier amplitude from 2.63 and 2.85 to 0.91, an improvement of about threefold. Applied to coarse finite-element simulations, the model reconstructs spectral content absent from the input, demonstrating transfer beyond filtered DNS. A conditional 3D-GAN trained on the same data failed to converge under Wasserstein training and is reported as a negative result. The main limitations are attenuation of the smallest resolved scales, periodic artefacts caused by the patch stride, and under-prediction of extreme velocity values.

Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Fluid Dynamics (physics.flu-dyn)

arXiv:2507.22082 [cs.LG]

(or arXiv:2507.22082v2 [cs.LG] for this version)

https://doi.org/10.48550/arXiv.2507.22082 Focus to learn more arXiv-issued DOI via DataCite

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.