Biomass gasification performance prediction based on mechanism simulation and data-driven

Significant nonlinearities in the biomass gasification process limit the ability of single-mechanism models and conventional experiments to make quick predictions under a variety of operating situations. This study suggests a paradigm for predicting biomass gasification performance that combines interpretable machine learning with Aspen Plus mechanistic simulation. By combining data from literature searches and independent Aspen simulations, a comprehensive database containing 1273 samples was constructed, covering 28 types of biomass feedstocks and typical operating conditions. After comparing the performance of Artificial Neural Network (ANN), Random Forest (RF), Extreme Gradient Boosting Regressor (XGBR), and Light Gradient Boosting Machine (LGBM) models, it was found that nonlinear models significantly outperformed linear benchmarks. Further optimization using Bayesian optimization resulted in R2 values above 0.96 for all eight target test sets using XGBR. SHapley Additive exPlanations (SHAP) was used to interpret the model predictions. External literature validation showed that the model has good generalization capabilities across different feedstocks, with a maximum deviation of approximately 5.6% for H2 and a maximum relative deviation of approximately 8% for LHV, demonstrating overall good agreement with trends. In summary, the mechanism-data fusion framework proposed in this study can achieve high-precision and interpretable prediction of multiple biomass gasification indicators, providing reliable support for gasification process optimization and downstream synthesis suitability assessment.