Principal component-based approach for kinetic model optimization

Kinetic models are essential for describing combustion chemistry and play a central role in developing cleaner combustion technologies. Their predictive accuracy is crucial for performing reliable combustion simulations, which support the design and optimization of these application systems. Model parameter optimization is often employed to improve the predictive accuracy of the kinetic models. For certain fuels like ammonia, where key reaction sensitivities span a broad range of temperatures, optimizing all three Arrhenius parameters (the pre-exponential factor A , the temperature exponent n , and the activation energy E ) yields significantly higher predictive accuracy than conventional A -factor optimization, owing to the additional degrees of freedom. However, the inherent correlations among the Arrhenius parameters pose challenges to the accuracy of response surface modeling, a commonly used strategy for efficient optimization. We present a novel optimization approach involving the projection of Arrhenius parameters into the uncorrelated principal component (PC) space. An artificial neural network-based response surface is employed. A significant improvement in the accuracy of the response surface model is observed when using PCs instead of the Arrhenius parameters. The objective function uses the curve-matching score, which quantifies both local and global agreement between experimental data and model predictions. Our optimization approach is applied to the NH 3 /NO combustion system as a test case, yielding an optimized thermal DeNO x model that exhibits a significantly improved predictive accuracy. A comparison between Arrhenius-based and PC-based optimization revealed superior performance with the PC-based approach. The optimized results allowed us to improve the understanding of key reactions in NH 3 combustion, including , , and , which sensitively impact the NH 3 /NO combustion process. Novelty and significance statement This study introduces a novel principal component-based optimization methodology for refining kinetic models, fundamentally distinct from traditional Arrhenius parameter optimization. By transforming correlated Arrhenius parameters into an orthogonal principal component space, this approach enhances the accuracy of surrogate response surfaces, resulting in an optimized model with superior predictive performance. Additionally, the novel integration of an ANN-based response surface and a CM score-based objective function in our framework addresses shortcomings of conventional techniques, marking a significant advancement in kinetic model optimization. Its successful application to thermal DeNO x chemistry as a practical demonstration case underscores its potential as a robust and efficient framework for future kinetic model optimization, ultimately contributing to improved combustion simulations and the advancement of next-generation combustion technologies.