Physics-informed machine learning methods for biomass gasification modeling by considering monotonic relationships

Area of Science:

• Chemical Engineering
• Computational Science
• Artificial Intelligence

Background:

• Machine learning (ML) shows promise in biomass gasification modeling but often lacks physical interpretability with limited data.
• Existing ML models struggle with generalization and ensuring physically plausible predictions in complex processes.

Purpose of the Study:

• To develop a physics-informed neural network (PINN) for predicting biomass gasification products (N2, H2, CO, CO2, CH4).
• To enhance the physical interpretability and generalizability of ML models in biomass gasification.

Main Methods:

• Developed a PINN that integrates regression, structure, and physical monotonicity constraints into the loss function.
• Compared PINN performance against five other ML methods using random sample classifications.

Main Results:

• PINN models achieved superior prediction capabilities with an average test R² of 0.91-0.97.
• The developed PINN models demonstrated correct monotonicity even with out-of-distribution data.
• PINN significantly outperformed other ML methods in prediction accuracy.

Conclusions:

• Physics-informed machine learning offers a pathway to physically feasible and interpretable biomass gasification models.
• PINN enhances model generalizability and reliability by embedding physical mechanisms.
• This approach addresses the interpretability limitations of conventional ML in scientific modeling.