Propagation of kinetic uncertainty through surrogate subspace in combustion simulations

In combustion simulations with uncertainty quantification, a large number of samples from the high-dimensional uncertainty parameter space are needed to propagate the kinetic uncertainty to global combustion characteristics such as the ignition delay time. Recognizing that it is computationally challenging to perform so many individual turbulent combustion simulations, a novel approach is proposed to propagate the kinetic uncertainty through a low-rank surrogate subspace as a replacement for the entire parameter space to reproduce the uncertainty on the target quantity. Specifically, the original parameter uncertainty hypercube is projected onto the surrogate subspace via a transform matrix, which is optimized using the generic algorithm, such that the projected samples in the surrogate subspace can reproduce the uncertainty of the target. We demonstrate the constructions of the surrogate subspaces for a detailed hydrogen mechanism, in which artificial neural network is employed as the response surface to accelerate the uncertainty propagate from the reaction rates to the target quantity. Then the probability distribution function of the above quantities is specified as the objective functions for the optimization. The surrogate subspace is optimized by specifying the ignition delay time, laminar flame speed and the extinction time in perfect stirred reactor as the target quantities, and is validated against the extinction strain rate in counterflow diffusion flame.