Reconstruction of Lean Hydrogen/Air Turbulent Boundary Layer Flames Using Generative Deep Learning

This study investigates the potential of a generative deep learning framework based on conditional rectified flow to reconstruct turbulent reactive flows. Through pre-training, the framework learns the probabilistic transport path from a Gaussian noise distribution to the two-dimensional slice distributions of velocity, temperature, and species mass fractions in near-wall turbulent combustion. After pre-training, the framework aligns the generation process with various distorted observations through iterative optimization sampling. This enables flexible multi-task reconstruction at different wall-normal positions without retraining. Evaluated against direct numerical simulation (DNS) data, the framework achieves structural similarity (SSIM) above 0.9 for sparse reconstruction tasks with sparsity greater than 5%, for denoising tasks with up to 15% Gaussian noise, and for super-resolution tasks with up to four-fold downsampling. Under mixed-distortion conditions combining noise and low-resolution observations, the framework effectively restores flow features and flame structures. It also demonstrates a strong generalization capability to unseen wall-normal positions, outperforming the U-Net model at far-wall locations with a root mean square error (RMSE) below 1.2 m/s for streamwise velocity and SSIM above 0.95. Additionally, the framework substantially reduces computational cost compared to DNS, offering new insights for combustion diagnostics and hydrogen system optimization.