Prediction enhancement through data augmentation techniques in AI-ML modelling of H₂ production from biomass...

Accurate prediction of hydrogen production in two-stage biomass gasification requires robust machine learning (ML) models, yet the scarcity of experimental data constrains model generalization. This study addresses data scarcity through integrated data augmentation strategies: permutation-invariant catalyst composition augmentation (PICCA) and Generative Adversarial Networks (GANs). A baseline literature dataset of 202 experimental samples was expanded to 1,450 datapoints across four hierarchical datasets with the combination of PICCA and GANs. Feature engineering via Spearman correlation eliminated redundancies, six regression ML models (Lasso, Ridge, Support Vector, Gaussian Process, Random Forest, and Gradient Boosting regressors) were trained using 5-fold cross-validation with randomized grid search hyperparameter optimization. Gradient Boosting Regression Trees (GBRT) achieved superior performance on the GAN + PICCA dataset (R² = 0.94, RMSE = 3.11, AAD = 2.30), representing a 43% accuracy boost over baseline dataset. Model interpretability via SHAP (i.e., Beeswarm & Waterfall) and partial dependence analysis revealed biomass type, bed material type & composition, and reformer & gasifier temperature as primary drivers of hydrogen yield. Results demonstrate that combined data augmentation strategies significantly enhance ML model predictive accuracy, enabling reliable integrated process and catalyst-driven optimization of biomass gasification for sustainable hydrogen production.