Accelerate global sensitivity analysis using artificial neural network algorithm: Case studies for combustion kinetic model

Global sensitivity and uncertainty analyses have attracted more and more attention in recent combustion kinetic studies. However, the high computational cost hinders their application in complex kinetic models. In order to accelerate the convergence speed, the artificial neural networks (ANN) methodology is applied into two widely used quantitative sensitivity analysis methods in the present work, i.e. the Sobol' sensitivity estimation and the random sampling high dimensional model representations (RS-HDMR). An ANN is constructed and trained using original model samples, which can then be used as a surrogate model to generate numerous samples for a global sensitivity analysis with Sobol' sensitivity estimation or RS-HDMR. It is shown that the ANN greatly reduces computational costs for estimating Sobol' sensitivity indices. The performances of the proposed ANN based HDMR method (ANN-HDMR) have been tested by a widely used analytical function (Sobol' g-function) and two practical models in combustion (master equation kinetic model and reaction kinetic model). The results show that the ANN-HDMR only needs a few tenths of original samples in the sensitivity analysis of master equation kinetic model and premixed H2/O2 ignition model. The ANN-HDMR is a kind of double-layer surrogate model which couples the advantage of ANN for fast convergence and RS-HDMR for direct sensitivity indices calculation, and thus it exhibits better performance in convergence and stability comparing with the commonly used RS-HDMR.