Search for a command to run...
To enhance the aerodynamic performance of organic Rankine cycle (ORC) turbines under increasing energy demands, surrogate-based optimization was applied to a 100 kW ORC turbine rotor. Four representative surrogate models—a radial basis neural network (RBNN), Kriging, response surface approximation (RSA), and a PRESS-based weighted (PBW) ensemble—were comparatively evaluated under identical numerical conditions. Independent optimizations of the first- and second-stage rotors enabled an examination of how different design variable space characteristics influenced surrogate predictive behavior. A fractional factorial sampling strategy was used to construct the training dataset, and learning curve analysis was conducted to assess sample size adequacy. Sensitivity estimation revealed distinct response surface characteristics between stages, allowing the interpretation of variations in surrogate stability. In both stages, geometric modifications were primarily concentrated near the outlet blade angle, identified as a dominant variable influencing efficiency. CFD validation confirmed that surrogate-based exploration successfully identified improved rotor geometries. Flow-field analysis indicated reduced entropy generation near the trailing edge region, suggesting the mitigation of aerodynamic losses. The results demonstrate that surrogate-based optimization can reliably improve turbine performance within a bounded design space, while the relative effectiveness of surrogate models depends on the sensitivity structure of the underlying problem.