Advancing clean combustible syngas production: an interpretable machine learning framework for biomass and plastic co-gasification

Abstract

Biomass–plastic co-gasification is an effective approach for producing clean and renewable energy and achieving sustainable development. However, the thermochemical reactions involved are complex, and the components of the gaseous products strongly depend on feedstock properties and operating conditions. Conducting conventional co-gasification experiments to produce syngas with high yield and desirable quality is time-consuming and labor-intensive. In this context, machine learning algorithms were employed to predict the production of syngas components (CO₂, CO, H₂, CH₄, hydrocarbons of C₂–C₄ (C_nH_m), and hydrogen/carbon monoxide ratio (H₂/CO)). Among the tested models, the category boosting (CatBoost) algorithm provided the best prediction performance for all output variables with coefficient of determination (R²) values of 0.80–0.94 and root mean square error (RMSE) values of 2.46–6.99 on the test set. Shapley additive explanations (SHAP) analysis indicated that temperature, steam/fuel ratio (S/F), biomass fixed carbon content, biomass proportion in feedstock, plastic hydrogen and oxygen content, and ash content are the most important factors influencing the yields of syngas components. High temperature promotes the conversion of CH₄ and CO₂ to H₂ and CO, while a higher S/F suppresses the formation of H₂ and CO. A higher biomass proportion increases CO₂ yield but decreases C_nH_m yield and the H₂/CO. The ash content of plastics serves as a valuable and representative proxy for both volatile matter and fixed carbon. These findings help deepen the understanding of key driving mechanisms in the gasification process, guide experimental design and process optimization, and enable targeted regulation of product yields, thereby improving syngas quality and utilization efficiency.

Graphical abstract