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

Area of Science:

• Materials Science
• Data Science
• Chemical Engineering

Background:

• Biochar properties depend on feedstock, modification, and pyrolysis, making engineered biochar design complex.
• Predicting biochar properties requires robust methods to handle data variability and interdependencies.

Purpose of the Study:

• To develop and evaluate data generative models for predicting biochar properties.
• To identify the most effective model for synthesizing high-fidelity biochar property data.

Main Methods:

• Four models were developed: Tabular Generative Adversarial Network (TGAN), Conditional Tabular Generative Adversarial Network (CTGAN), Tabular Variational Autoencoder (TVAE), and Synthpop.
• Models were trained and evaluated on imputed biochar property datasets (n=461).
• Model performance was assessed using distribution similarity, correlation preservation, and experimental validation.

Main Results:

• Synthpop demonstrated superior synthetic data quality, accurately capturing various distribution patterns.
• Synthpop achieved high distribution similarity (0.97), KSComplement (0.98), and TVComplement (0.95).
• Experimental validation showed low relative errors (<5%) for key biochar properties using Synthpop-generated data.

Conclusions:

• Synthpop is a reliable model for synthesizing pyrolyzed biochar properties.
• The developed framework enables rapid screening of application-specific biochar.
• Accurate synthetic data generation can accelerate biochar research and development.