Abstract Supervised machine learning is used to classify a continuous and deterministic design space into a nondominated Pareto frontier and dominated design points. The effect of the initial training data quantity on the Pareto frontier and output parameter sensitivity is explored. The study is performed with the optimization of a subsonic small-scale cavity-stabilized combustor. A 3D geometry is created and parameterized using computer aided design (CAD) that is combined with a software for meshing, which automatically transfers grids and boundary conditions to the solver and postprocessing tool. Steady, compressible three-dimensional simulations are conducted employing a multiphase Realizable k–ε Reynolds-averaged Navier–Stokes (RANS) physics with an adiabatic flamelet progress variable (FPV) model. The near-wall turbulence modeling is computed with scalable wall functions (SWFs). For each computational fluid dynamics (CFD) simulation, four levels of adaptive mesh refinement (AMR) are utilized on the original cut-cell grid. There are 15 geometrical input parameters and three output parameters, viz., a pattern factor proxy, a combustion efficiency proxy, and total pressure loss (TPL). Three times the number of input parameters plus one (48) is necessary to yield an optimization independent of the initial sampling. This conclusion is drawn by examining and comparing the Pareto frontiers and global sensitivities. However, the latter provides a better metric. The relative influence of the input parameters on the outputs is assessed by Spearman's order-rank correlation and an active subspace analysis. Some persistent geometric features for nondominated designs are also discussed.