Cross-Source Validation of Pressure-Ratio-Driven Surrogate Models for Predicting Steam Ejector Entrainment Ratio

Pressure-ratio-based surrogates provide compact estimates of steam-ejector entrainment ratios (ERs), but internal accuracy does not establish experimental predictive reliability. This study evaluated four original regression configurations using 111 CFD development records and 18 experimental records representing 17 unique conditions in the same ejector geometry. Leave-one-out, input-pair-grouped and subset-holdout evaluations were combined with independent H1 testing, input-coverage analysis and TreeSHAP. Gaussian process regression and XGBoost achieved leave-one-out R2 of 0.9271 and 0.9272, respectively, yet all four configurations yielded negative H1 R2, including for the 16 records inside the development-input convex hull. Adding Ridge and support vector regression with development-only hyperparameter selection and nested grouped validation still yielded negative H1 R2 across all six selected configurations. All four original models overpredicted ER at every one of the 17 conditions, with median errors 26.3–42.2 times the reported ±3.5% measurement-uncertainty band; removing any single condition preserved negative R2. TreeSHAP assigned 77.9% of the summed mean absolute attribution to compression ratio, without establishing physical causality. These findings identify a gap between internal accuracy and experimental transfer and support installation-specific validation before monitoring applications.