Uncertainty-based weight determination for surrogate optimization

Abstract

Fuel surrogates are simplified models that mimic the combustion characteristics of very complex real transportation fuels and enable a detailed description of the computational system for the targeted real fuels. Current efforts in surrogate development focus on matching multiple target properties of the real fuel using numerical optimization of species compositions. A way to solve the multi-objective optimization problem is to employ the weighted-sum approach with arbitrarily assigned weights. In this paper, we propose a novel approach to reduce such arbitrariness by incorporating physical information into the surrogate optimization process, leveraging uncertainties from experimental measurements and mixture property predictions to determine the weight of each target property. The underlying principle of this approach is assigning low weight for properties with high uncertainty since a tight emulation of that property is not necessary. We propose formulations that convert relative uncertainties of each target property into weights for multi-objective surrogate optimization, which penalize target properties with high uncertainties. The method is flexible enough to accommodate not only multiple uncertainty sources, but also users’ preferences and the sensitivity of weights to uncertainty variations can be readily adjusted. The proposed method is applied to formulate surrogate mixtures for three reference target jet fuels (Jet-A POSF-10325, JP-8 POSF-10264, JP-5 POSF-10289), which have considerably different hydrocarbon compositions and properties. The results show that the surrogates developed by the uncertainty-based weight method emulate the target properties of fuels very closely, and they are able to capture the compositional characteristics of the target fuels as well.