Data Quality and Generalization Reliability of Small-Sample Machine-Learning Surrogates: A CFD-Based Savonius Wind-Turbine Case Study

Machine-learning surrogates are widely used to reduce the cost of computational fluid dynamics (CFD), but reliable performance can be difficult to establish when only a small and non-uniformly sampled design set is available. This study examines data quality and generalization reliability using 33 CFD configurations generated for a Savonius wind-turbine deflector study. Three geometric variables where were deflector number, normalized length, and angle, were used to predict driving-side velocity (Vp), returning-side velocity (Vn), and their difference (ΔV). Six surrogate models were compared using leave-one-out cross-validation (LOOCV), and the best response-specific models were then subjected to training-size, design-set transfer, predictive-uncertainty, noise, and outlier analyses. Gaussian-process regression predicted Vp well under LOOCV (RMSE 0.226 m/s, R² 0.858), whereas Vn remained weakly learnable (best R² 0.173) and ΔV showed intermediate performance (R² 0.389). Increasing the Vp training set from 8 to 28 configurations reduced mean matched-test RMSE by 55.7%. However, strong within-dataset accuracy did not guarantee transfer between the factorial and supplementary design sets: GPR Vp performance fell to R² = −0.125 in the factorial-to-supplementary direction. Predictive uncertainty ranked errors under this shift but was poorly calibrated in one direction. Controlled label noise and outlier contamination further degraded the otherwise learnable Vp surrogate. The results show that small-sample surrogate reliability depends jointly on data quantity, target learnability, DOE coverage, corruption sensitivity, and uncertainty calibration, not on cross-validation accuracy alone.