Global sensitivity analysis (GSA) can support data-efficient optimization by identifying the most influential inputs and enabling dimensionality reduction when high-fidelity data are scarce. This work combines variance-based (Sobol’) GSA with maximum-projection experimental designs and Bayesian optimization to detect input anisotropy and focus a limited simulation budget on an active low-dimensional subspace. The approach is demonstrated on a computationally expensive CFD model of a radial turbine relevant to polymer electrolyte membrane fuel cell propulsion systems. An initial maximum-projection design is generated in a 10-dimensional parameter space and then sequentially enriched using Bayesian optimization with a Gaussian-process surrogate and an upper-confidence-bound acquisition function. In parallel, polynomial chaos expansion surrogates are trained to compute sensitivity metrics. To ensure robust surrogate modeling of the bounded turbine efficiency, the response is transformed to an unconstrained space via cumulative-distribution normalization followed by a logit mapping. Sensitivity results are used to identify active parameters and continue Bayesian optimization in the reduced input space. With only a few hundred simulations, the proposed workflow increases turbine efficiency by 6% compared to the best design in the initial experimental design.