Description
The ever-increasing focus of policy-makers on environmental issues are pushing the combustion community towards making combustion cleaner by optimizing the combustion equipment in order to reduce emissions, improve efficiency and satisfy the increasing energy demand. A major part of this involves advancing modelling capabilities of these complex combustion systems, which is a combination of computational fluid dynamics with detailed chemical kinetic models. A chemical kinetic model comprises of a series of elementary reactions with corresponding kinetic rate parameters and species thermodynamic and transport data. The predictive capability of these models depends on the accuracy to which individual chemical reaction rates, thermodynamic and transport parameters are known. A minor fraction of the rate constants and thermodynamic properties in the widely used kinetic mechanisms are experimentally derived or theoretically calculated. The remaining are approximated using rate rules and group additivity methods respectively for rate constants and thermodynamic properties. Recent works have highlighted the need for error checking when preparing such models using the approximations, but a useful community tool to perform such analysis is missing. In the initial part of this work, we developed a simple online tool to screen chemical kinetic mechanisms for bimolecular reactions exceeding collision limits. Furthermore, issues related to unphysically fast time scales can remain an issue even if all bimolecular reactions are within collision limits. Therefore, we also presented a procedure to screen ultra-fast reaction time scales using computational singular perturbation (CSP). The screening of kinetic models is a necessary condition, however, not a sufficient one. Therefore, exploring new approaches for the simulation of complex chemically reacting systems are needed. This work focuses on developing new methods for estimating thermodynamic data efficiently and accurately, thereby increasing the compliance of forth-mentioned screening. Machine Learning (ML) has been increasingly becoming a tool of choice for regression, replacing traditional function fittings. Group additivity incorporates simple functions and derive constants with a certain existing data and use these functions to estimate the unknown values. ML algorithms does the same without fixing a specific function there by letting algorithm to learn the non-linearity from the training data itself. With the new data coming in with time, ML algorithms learn better and improves over time, whereas this need not necessarily happen with traditional methods. In the first part of the study, data for standard enthalpy is collected from the literature sources and ML models are built on these databases. Two different models were built and studied for a straight-chain species and cyclic species dataset. Molecular descriptors are used as the datasets collected from literature are small for using any sparse representations as input. As expected, we observed a good improvement above group additivity method for these ML models. The improvement is observed to be more significant for cyclic species. With the motivation of ML models showing benefit over the group additivity method, a step further was taken. A homogenous and accurate dataset is necessary for building a ML model that can be used for generating the thermodynamic data for kinetic models. With this in mind, an accurate database for thermodynamic data is built from ab-intio calculations. The species in the dataset are taken from a detailed and well established mechanism to cover all the species in a typical kinetic mechanism. The calculations are performed at a high level of accuracy, in comparison to other similar datasets in literature. In the later part of this work, the dataset developed using ab-inito calculations is used for developing ML models. Unlike the ML models built from the literature datasets, this database consists of all the thermodynamic data required for kinetic models viz. standard enthalpy and standard entropy and heat capacity at 300 K and higher temperatures. To numerically mimic real gasoline fuel reactivity, surrogates are proposed to facilitate advanced engine design and predict emissions by chemical kinetic modelling. However, chemical kinetic models could not always accurately predict non-regular emissions, e.g. aldehydes, ketones and unsaturated hydrocarbons, which are important air pollutants. Therefore, we propose to use machine-learning algorithms directly to achieve better predictions, circumventing the kinetic models. Combustion chemistry of fuels constituting of 10 neat fuels, 6 primary reference fuels (PRF) and 6 FGX surrogates were tested in a jet stirred reactor. Experimental data were collected in the same setup to maintain data uniformity and consistency. Measured species profiles of methane, ethylene, propylene, hydrogen, carbon monoxide and carbon dioxide are used for machine-learning model development. The model considers both chemical effects and physical conditions. Chemical effects are described as different functional groups, viz. primary, secondary, tertiary, and quaternary carbons in molecular structures, and physical conditions as temperature. Both the Machine-learning models used in this study showed a good prediction accuracy. By expanding the experimental database, machine-learning models can be further applied to many other hydrocarbons in future work, for the direct predictions.
Date made available | 2022 |
---|---|
Publisher | KAUST Research Repository |