Application of the neural surrogate optimization method to refine the skeletal chemical-kinetic mechanism of aviation kerosene combustion
V.V. Matyushkov1, A.G. Shmakov1,2, S.A. Trubachev1
1Voevodsky Institute of Chemical Kinetics and Combustion, Siberian Branch of the Russian Academy of Sciences, Novosibirsk, Russia
2Kutateladze Institute of Thermophysics, Novosibirsk, Russia
Keywords: neural surrogate, optimization, aviation kerosene, skeletal chemical-kinetic mechanism, laminar flame
Abstract
The work uses the neurosurrogate optimization method to refine the skeletal chemical-kinetic mechanism of KCPsur, which predicts the combustion rate of the model 4-component surrogate aviation fuel SU4 (a mixture of n-decane, isocetane, methylcyclohexane and tetralin). Multilayer perceptron trained on 34 variations of 5 pre-exponential multipliers of rate constants of chemical reactions that have the highest sensitivity coefficient of laminar combustion rate (flow rate) A-factor sensitivity) at 27 control points, was used instead of time-consuming calculations of the speed of propagation of laminar flames in the CHEMKIN software package. To optimize the rate constants of the selected reactions responsible for the conversion of tetralin, the gradient descent method was used, which made it possible to obtain a refined KCPsur_neural mechanism. The new chemical-kinetic model more accurately describes the laminar flame propagation velocity (LBV) of kerosene-air mixtures under lean and near-stoichiometric conditions (φ = 0.70 - 1.05) at pressures up to 8 atm, surpassing the accuracy of the original KCPsur and, in some cases, the detailed CRECK mechanism. An important feature of KCPsur_neural is its ability to correctly describe the concentration profiles of a number of important intermediates in the flame of a SU4/O2/Ar mixture, which indicates a weak influence of the rate constants of refined reactions on the structure of the studied flames. This opens up the possibility of using two independent neurosurrogate models in the future to optimize fuel combustion mechanisms: one to optimize them with respect to LBV, the other to describe the flame structure with the formation of separate training samples of a smaller total dimension. It is shown that the proposed approach speeds up the optimization of reaction rate constants by approximately 106 times when using a limited training set. The results demonstrate the high effectiveness of using machine learning to overcome the rigidity of reduced combustion models and expand their applicability to conditions relevant to aircraft engine combustion chambers.
|
