Data-driven modeling of near-wall turbulence using β-variational autoencoder, transformers, and adversarial loss
Published in 11th International Symposium on Turbulence Heat and Mass Transfer (THMT-25), 2025
Abstract
This work presents a machine learning framework for data-driven modeling of near-wall turbulence, combining a β-Variational Autoencoder-Generative Adversarial Network (β-VAE-GAN) to extract dominant flow features into a low-dimensional latent space with a decoder-only transformer modeling their temporal evolution. Validated on 2D DNS snapshots of a minimal turbulent channel flow at (Re_\tau = 200), the framework accurately predicts latent space dynamics, preserves chaotic characteristics, and enables reconstruction of low-velocity streaks and dominant momentum transport mechanisms within one Lyapunov time.
Recommended citation: Tonioni, N., Umair, M., Agostini, L., Kerhervé, F., Cordier, L., & Vinuesa, R. (2025). Data-driven modeling of near-wall turbulence using β-variational autoencoder, transformers, and adversarial loss. THMT-25, Tokyo, Japan.
Download Paper