Tabular generative modeling framework for multi-property data synthesis of pyrolyzed biochar

Biochar properties are governed by trade-offs among feedstock types, modification methods, and pyrolysis conditions, complicating the design of engineered biochar for specific applications. In this study, four data generative models, including Tabular Generative Adversarial Network (TGAN), Conditional Tabular Generative Adversarial Network (CTGAN), Tabular Variational Autoencoder (TVAE), and statistical Synthpop, were developed to predict biochar properties using MultipleMICE- and MissForest-imputed reference datasets (n = 461). The Synthpop model outperformed others in synthetic data quality, achieving high distribution similarity score (0.97) by accurately capturing multimodal, normal, and skewed distribution patterns for both continuous features (KSComplement = 0.98) and categorical variables (TVComplement = 0.95). Correlation trends evaluation indicated that TVAE and Synthpop models effectively preserved inter-feature dependencies between synthetic and original real data. Experimental validation with bagasse- and bamboo-derived biochar confirmed the reliability of Synthpop model in generating high fidelity data for biochar properties such as surface morphology, elemental composition, and surface functional groups. Specifically, up to 70.8 % of features for bagasse biochar and 66.7 % of features for bamboo biochar exhibited low relative errors (<5 %). Overall, this work introduced an applicable framework and a reliable Synthpop model for synthesizing pyrolyzed biochar properties, providing insights into rapid screening of application-specific biochar.