Predictions of instantaneous temperature fields in jet-in-hot-coflow flames using a multi-scale U-Net model

• DOI: 10.1016/j.proci.2024.105330
• Publication Date: 2024-07-05T17:52:36Z
• Authors: Jordan A.C. Kildare; Wai Tong Chung; Michael J. Evans; Zhao F. Tian; Paul R. Medwell; Matthias Ihme
• Keywords: Jet fuel

A multi-scale U-Net machine learning (ML) model is developed to assess its validity as a surrogate for non-intrusive flame temperature measurement in jet-in-hot-coflow (JHC) flames. Inputs the are simultaneous hydroxyl (OH) and formaldehyde (CH2O) planar laser-induced fluorescence (PLIF) measurements, with target fields derived from Rayleigh scattering measurements. Coflow oxygen (O2) concentration, jet Reynolds number, coflow temperature, fuel inputs were considered dataset, resulting 33 unique conditions, ∼17,000 training images. The ML reconstructs instantaneous images within an absolute error of 7 ± 10% measured values initial natural gas/ethylene model. Transfer was employed accelerate across conditions different fuels, mean reconstruction 9% models ethanol dimethyl ether. fourth including data all four fuels trained using transfer learning, 6 error. Multi-layer perceptron (MLP) classification applied deepest layers four-fuel identify whether physical chemical features being encoded. Experimentally verifiable found be encoded latent space model, demonstrating that it not solely image regression tool, but capable predicting type, downstream axial location, O2 concentration. These findings indicate only well suited predictions, can provide scientific insight well.