A new framework of global sensitivity analysis for the chemical kinetic model using PSO-BPNN

Global sensitivity analysis is a tool that primarily focuses on identifying the effects of uncertain input variables on the output and has been investigated widely in chemical kinetic studies. Conventional variance-based methods, such as Sobol' sensitivity estimation and high dimensional model representation (HDMR) methods, are computationally expensive. To accelerate global sensitivity analysis, a new framework that combines a variance-based (Wu's method) and two ANN-based sensitivity analysis methods (Weights and PaD) was proposed. In this framework, a back-propagation neural network (BPNN) methodology was applied, which was optimized by a particle swarm optimization (PSO) algorithm and trained with original samples. The Wu's method and Weights and PaD methods were employed to calculate sensitivity indices based on a well-trained PSO-BPNN. The convergence and accuracy of the new framework were compared with previous methods using a standard test case (Sobol' g-function) and a methane reaction kinetic model. The results showed that the new framework can greatly reduce the computational cost by two orders of magnitude, as well as guaranteeing accuracy. To take maximum advantage of the new framework, a four-step process combining the advantages of each method was proposed and applied to estimate the sensitivity indices of a C2H4 ignition model. The sensitivity indices of the more complex model could be implemented easily with good accuracy when the four-step process is followed