Large eddy simulations of ammonia-hydrogen jet flames at elevated pressure using principal component analysis and deep neural networks

Abstract

The combustion of ammonia/hydrogen is currently gaining importance in the power generation sector as an alternative for hydrocarbon fuels and improved fundamental insights will facilitate its application. To investigate the complex interactions between turbulence and chemistry for ammonia-hydrogen jet flames under high-pressure conditions, large eddy simulation (LES) computations are conducted using the PC-transport model, which is based on Principal Component Analysis (PCA), coupled with nonlinear regression that utilizes deep neural networks (DNN) to enhance the size-reduction potential of PCA. Classical statistics-based nonlinear regression methods are inefficient at fitting highly nonlinear manifolds and when large data sets are involved. These two drawbacks can be overcome by utilizing DNN tools. Several DNN architectures composed of fully connected layers of different depths and widths, batch normalization, and various activation functions coupled with various loss functions (mean squared error, absolute error, and R2) are explored to find an optimal fit to the thermo-chemical state-space manifold. The ability to achieve highly accurate mapping through DNN-based nonlinear regression with an R2-score >0.99 is shown by employing a single graphical processing unit (Nvidia RTX 3090). Furthermore, the proposed PC-DNN approach is extended to include differential diffusion based on a rotation technique and utilization of the mixture-averaged transport model for the training data set. To demonstrate the potential of the PC-DNN approach in modeling turbulent non-premixed combustion, LES results are compared with the recent Raman/Rayleigh scattering measurements that were obtained at the KAUST high-pressure combustion duct (HPCD). Results show that the PC-DNN approach is able to capture key flame characteristics with reasonable accuracy using only two principal components. The inclusion of differential diffusion leads to improved predictions, although some discrepancies are observed in fuel-lean regions.