Biomass gasification modeling based on physics-informed neural network with constrained particle swarm optimization

Significant interest has emerged in using machine learning (ML) methods to predict biomass gasification performance. However, insufficient experimental data often hinders ML models' ability to accurately characterize the gasification process. Therefore, this study proposed a novel hard-constrained physics-informed neural network (HC-PINN) to predict biomass gasification syngas components. This method employs a hard-constrained learning approach and a constrained particle swarm optimization algorithm to ensure that the constructed model fully obeys the given monotonicity constraints and aligns with experimental data. A comprehensive dataset collected from the existing articles was utilized to model the biomass gasification process by the HC-PINN and six advanced ML approaches. The results show that the HC-PINN model demonstrates impressive performance in test sessions, achieving an R2 value greater than 0.95 for Case 1 (where the training and testing sets are randomly assigned) and over 0.90 for Case 2 (where training and testing sets do not contain the same biomass feedstock). Specifically, the HC-PINN model can strictly adhere to the prior physical monotonicity even for the feedstocks that do not appear in the training samples, with the physical consistency degree equal to 1, reflecting better interpretability and generalization than the six state-of-the-art ML models.