An experimental design algorithm applied to study of methyl radical and H atom attack on normal butane

David A. Sheen, Iftikhar A. Awan, and Jeffrey A. Manion

Chemical Sciences Division, National Institute of Standards and Technology, Gaithersburg, MD 20899


We present a theoretical study of H and CH3 attack on normal butane (n-C4H10). The oxidation of hydrocarbon fuels proceeds by means of the attack of small radicals such as H and CH3 on large molecules. These radicals abstract hydrogen atoms from the large molecules, which then usually proceed by β-scission to form C2H4 and C3H6.† C2H4 is significantly more reactive than C3H6, and so the overall reactivity of a combustion system will depend on the relative amounts of these intermediates that are produced. A quantitative understanding of the radical attack process is critical to the development of chemical models for the oxidation of hydrocarbons. Furthermore, the measurement uncertainty will affect the rate parameter estimates, and it is useful to know how these affect the uncertainty in predicted global properties such as ignition delay time.

In this work, we propose an experimental design algorithm that will be applied to the problem of measuring H and CH3 attack rates on n-C4H10. We generate a set of proposed experiments covering a wide range of initial reactant concentrations and temperatures, and put forward a candidate model to simulate these experiments, in this case the Jet Surrogate Fuel model. We then use a machine-learning algorithm to identify the best subset of experiments to perform. In order to test the machine algorithm, we compare its performance against an expert-recommended set of experimental measurements.† We find that the machine-generated experimental set performs better than the expert-generated experimental set.† The machine learning algorithm is therefore a suitable surrogate for an expertís evaluation of a set of experiments, and can be applied to many other database analysis and constraint problems. In addition, when the uncertainty in the shock-tube measurements is propagated into predicted ignition delay times, we find that the uncertainty is substantially reduced and that the modelís ability to reproduce the experimental measurements is greatly improved.