Abstract
The paper presents a data based surrogate model of the chemical kinetics of hydrogen oxidation by air using recurrent and feed-forward neural networks. The work aims at the application of surrogate models in computational fluid dynamics simulators, which are ubiquitous in the development and optimization of modern chemical technologies. The sensitivity of the results to the size of the data set and network parameters is analyzed. For a seven-component reaction mechanism at adiabatic conditions, a model trained on a sample of one million sets of initial conditions enables prediction of the dependence of concentrations and temperature on time with a standard deviation below 2% over 20 microsecond range. However, points with large deviations reaching 10% are also observed, mostly for minor components with low concentrations. The surrogate model is several times faster compared to the direct numerical solution of kinetic equations on the temporal grid. The computational performance strongly depends on the batch size and is sensitive to the hardware. The results of the work demonstrate a significant potential of machine learning methods for modeling chemical transformations in computational fluid dynamics solvers. Further improvement of the accuracy with a similar computational performance can be expected from: (a) separate models for short-time (that is, strongly non-equilibrium) and long-time (closer to the equilibrium) ranges; (b) repeated optimization of network parameters even with minor modifications of the reaction mechanism; (c) more versatile approaches to complying with the conservation laws (d) application of physics informed machine learning (e.g. of the models with additional physical and chemical constraints such as mass conservation).