Core Concepts
DDPM outperforms heteroscedastic models in predicting the distribution of solutions for airfoil flow simulations.
Abstract
The study explores using DDPMs to train an uncertainty-aware surrogate model for turbulence simulations around airfoils. Results show DDPMs accurately estimate simulation uncertainties compared to Bayesian neural networks and heteroscedastic models. The analysis reveals a correlation between uncertainty and flow separation, highlighting challenges in RANS simulations. Experiments demonstrate DDPM's superior accuracy in predicting expectation and standard deviation fields compared to heteroscedastic models across varying parameters. The study provides insights into modeling uncertainties in fluid dynamics using advanced deep learning techniques.
Stats
𝑅𝑒 ∈ {1.5×10^6, 3.5×10^6, 5.5×10^6, 7.5×10^6, 9.5×10^6}
𝝈y,𝑎 < 5 × 10^-3 and 𝝈y,𝑎 ≥ 5 × 10^-3 used for test dataset evaluation