Effect of parametric uncertainty in numerical simulations of a hydrogen-fueled flameless combustion furnace using dimensionality reduction and non-linear regression

The objective of this work is to assess the propagation of the uncertainty in the 2D RANS model of a semi- industrial furnace, due to 9 uncertain input parameters. To collect the data required for this statistical analysis, Proper Orthogonal Decomposition combined with Gaussian Process Regression (POD/GPR) was employed to generate a surrogate model of a 2-dimensional Reynolds-Averaged Navier-Stokes (RANS) simulation in a 9dimensional parameters space. The surrogate model is built by compressing the training data using POD, which reduces the dimensionality of the RANS response. GPR is then used for regression in the uncertain parameter space. We apply this methodology to a hydrogen-fueled semi-industrial furnace to assess the predictive uncertainty of the RANS simulations due to the uncertainty associated with turbulence, combustion and kinetics model input parameters. The results show that the POD/GPR surrogate model is able to accurately predict the temperature and species mass fractions in the furnace. We employ the surrogate to perform a global sensitivity analysis to determine the relative importance of the uncertain input parameters. It is found that the turbulence parameter C 2 " and the combustion model constant C mix hold the strongest influence on the variability of the model response in the flame region, while the parameter controlling uncertainty of the inlet air mass flow rate is the one contributing the most in the recirculating region of the furnace. This study demonstrates the effectiveness of the POD/GPR approach in quantifying uncertainty in combustion systems and provides valuable insights into the contribution of different parameters to the response variance.