Abstract
Particle image velocimetry (PIV) is a standard method for studying primary recirculation zones in trapped vortex combustors (TVCs), which can operate in an RQL configuration. However, its intrusive nature can disrupt the flow, flame, and equipment in compact combustors, leading to inaccuracies. As an alternative, we use deep learning models based on generative adversarial networks (GAN, a widely used approach) and vision transformers (ViT, a recently devised promising architecture) to estimate the position and overall structure of large-scale vortices from a non-invasively measured quantity, such as the planar laser-induced fluorescence (PLIF) of a species. These models are trained using datasets from large-eddy simulations (LES) of TVCs with information regarding all scalars constituting the state variable, with the addition of noise to mimic experimental data. Quantitative metrics such as relative errors and PDFs of velocity components and their orientation have been used to demonstrate that the ViT exhibits better performance than the GAN. Sensitivity to the type of noise added to simulation data during training is studied as well. The trained model is then used to infer velocity vectors from noisy OH-PLIF data. In the absence of ground truth for that case, qualitative observations reinforce our earlier notion of the superiority of ViT. Such models will facilitate intelligent data fusion and the development of digital twins of combustors.
