Machine-learning-assisted physicochemical optimization of an RP-3 surrogate fuel model

To improve the physicochemical fidelity of RP-3 surrogate fuels, this study proposes a machine-learning-assisted multi-objective optimization framework that jointly considers conventional physical properties and combustion-related characteristics. A four-component surrogate consisting of n-decane, n-dodecane, ethylcyclohexane, and p-xylene was investigated. Density, viscosity, molecular weight, and hydrogen-to-carbon ratio were selected as physical-property targets, while ignition delay time (IDT) and laminar flame speed (LFS) were selected as chemical-property targets. Dirichlet sampling was employed to generate compositions uniformly over the constrained mixture simplex, and machine-learning surrogate models were trained using Cantera-generated IDT and LFS datasets to reduce the computational cost of objective evaluation. The NSGA-II algorithm produced a Pareto front showing a clear trade-off between the physical-property and chemical-property errors. A representative compromise solution was selected using the normalized distance to the ideal point. The resulting mole fractions of n-decane, n-dodecane, ethylcyclohexane, and p-xylene were 0.038, 0.637, 0.264, and 0.061, respectively. After re-evaluation using the detailed chemical kinetic mechanism, the physical-property error decreased from 0.0355 to 0.0106, while the chemical-property error decreased from 0.0936 to 0.0665 relative to the baseline formulation, corresponding to reductions of 70.1% and 28.9%, respectively. Sensitivity analyses involving different physical-property weights, machine-learning models, and multi-objective optimization algorithms showed that the dominant composition trend remained consistent. These results demonstrate that the proposed framework provides a computationally efficient method for optimizing RP-3 surrogate fuels while explicitly accounting for the trade-off between physical properties and chemical characteristics within the investigated operating range.