Gaussian process kernel search applied to RON and S prediction
One factor in estimating the performance potential of a new automobile fuel prior to market introduction is understanding the fuel’s auto ignition propensity, historically qualified by its research octane number (RON), motor octane number (MON) and sensitivity (S = RON – MON). In order to estimate the octane numbers, researchers typically need multiple liters of the new fuel, potentially involving a prohibitive expense for exotic new fuels. This project attempts to predict the octane numbers using small sample volumes, less than 0.5 ml of fuel; this would mean lower costs in screening new fuels, leading to more potential fuels being considered. To predict the octane numbers, we build Gaussian process models. We implement a greedy computational kernel search to optimize the types of structures our Gaussian process can learn. Results suggest these approaches are both computationally effective and efficient.
(Bagels and refreshments will be served.)