Abstract
Reliable characterization of the upper and lower flammability limits of methane–hydrogen mixtures is imperative for the safe application of hydrogen-enriched fuel systems. However, experimental data under elevated-temperature and elevated-pressure conditions remain sparse and heterogeneous. The present study evaluates Random Forest, Gradient Boosting, and XGBoost models using literature-derived data and compares reference-based sampling, a physics-inspired feature derived from the Arrhenius equation and collision theory, and conditional tabular generative adversarial network (CTGAN) augmentation. Reference-based sampling has been demonstrated to reduce internal validation errors; however, unfiltered CTGAN-generated data has been shown to degrade predictive performance. Following the screening of the synthetic data, the XGBoost mean squared error (MSE) decreased from 10.369 to 4.510, and the mean absolute error (MAE) decreased from 1.342 to 0.853. The Random Forest model demonstrates relatively stable composition-dependent trends, while the XGBoost model exhibits a more pronounced response in hydrogen-rich mixtures. Predictions at 200 °C and 3 MPa remain exploratory due to the unavailability of direct experimental data covering the complete composition range. These findings underscore the significance of reference-based sampling and synthetic-data screening in the context of machine learning applied to sparse flammability-limit datasets.