Disentangled Representation Aided Physics-Informed Neural Network for Predicting Syngas Compositions of Biomass Gasification

Machine learning (ML) has been extensively studied and applied in the biomass gasification field currently. However, insufficient experimental data tends to cause a mismatch between the ML model and physical mechanism, particularly for the feedstocks that do not appear in the training data set, becoming a significant challenge in creating credible ML models for biomass gasification. Therefore, this study proposes a disentangled representation-aided physics-informed neural network method, briefly called DR-PINN, to predict biomass gasification syngas components. First, the DR-PINN extracts the latent variables to represent the feedstock properties through disentangled representation learning and generates synthetic samples in the gasification-related latent variable space to cover a full range of feedstock types. Then, DR-PINN employs inequality constraints to embed a priori monotonic relationships into the model training loss function. Finally, experimental and synthetic samples are simultaneously considered in the model training process to realize the synergy and complementarity of actual data information and existing physical knowledge using an evolutionary algorithm. As a result, DR-PINN shows good prediction performance (the feedstocks within the training data set: R2 ≈ 0.96, root-mean-square error (RMSE) ≈ 1.7; the feedstocks outside the training data set: R2 ≈ 0.81, RMSE ≈ 3). Moreover, even with the feedstocks outside the training data set, the DR-PINN model can strictly abide by the prior physical monotonic relationships, with the physical consistency degree equal to 1. Overall, the proposed DR-PINN demonstrates superior generalization and interpretability compared to other methods, such as RF, GBR, SVM, ANN, and PINN.