<p>Machine Learning Prediction of Higher Heating Value in Heterogeneous Biomass: An Evaluation of Synthetic Data Augmentation Strategies</p>

Machine learning and generative artificial intelligence are increasingly used to predict biomass properties in resource-constrained settings where experimental determination is impractical. Experimental determination of the higher heating value (HHV) via bomb calorimetry remains costly, time-intensive, and largely inaccessible in resource-constrained regions, motivating the development of data-driven alternatives for biomass feedstock screening. This study is the first to evaluate synthetic data augmentation for biomass HHV prediction, moving beyond the visual comparisons used in prior studies toward a quantitative assessment of synthetic data quality. This study evaluates the performance of five machine learning models and examines the impact of synthetic data augmentation on their predictive performance. The models include Support Vector Regression, Bayesian Ridge, Extremely Randomized Trees (Extra Trees), Gradient Boosting, and Extreme Gradient Boosting (XGBoost). All models are trained on a heterogeneous dataset of 315 biomass samples across six feedstock categories obtained from the literature. XGBoost achieved the highest test-set performance, with an R² of 0.81 and a root mean square error of 0.94 MJ/kg. The study also evaluates synthetic data generated for biomass HHV prediction using conditional Generative Adversarial Networks (cGAN), Gaussian Copula, and Bootstrap with noise, assessed across fidelity, utility, and privacy metrics. The cGAN demonstrated the most balanced performance, with a Hellinger distance of 0.17, a pairwise correlation difference of 0.09, and a memorization rate of 3.40%. The results establish that augmentation utility is model-dependent. The most notable gains were observed under cGAN augmentation for Gradient Boosting (ΔR² = +0.028) and XGBoost (ΔR² = +0.019), with XGBoost reaching an R² of 0.83. Extra Trees showed consistent improvement under Gaussian Copula augmentation, with the mean absolute error reduced from 0.93 to 0.90 MJ/kg. However, performance deteriorated across all models when synthetic sample sizes exceeded 25-100 samples.