Explorando a produção de hidrogênio a partir da co-gaseificação de biomassa–plástico: um estudo integrado de simulação e aprendizado de máquina

Abstract

Biomass–plastic co-gasification is a promising route for producing low-carbon syngas and hydrogen; however, its optimization is challenged by nonlinear interactions among temperature, equivalence ratio, and feed composition. In this work, an integrated approach combining steady-state thermochemical process modelling and machine learning was developed to predict and optimize the performance of biomass–plastic systems. A phenomenological model was implemented in the AVEVA PRO/II simulator, structured into drying, pyrolysis, gasification, and restricted chemical-equilibrium stages, using air as the gasifying agent. The model was validated against experimental literature data, reproducing the order of magnitude of the molar fractions of H₂, CO, CO₂, and CH₄ under different operating conditions. Based on this validated platform, a synthetic dataset of 3,000 simulations was generated via Latin Hypercube Sampling, covering five plastics (HDPE, PE, PP, PS, and PET), 35 lignocellulosic biomasses, and representative ranges of temperature and equivalence ratio. Extreme Gradient Boosting (XGBoost) models were then trained to predict syngas composition (H₂, CO, and CO₂), total gas yield, lower heating value, and the H₂/CO ratio. The split into training, validation, and test sets was assessed using distance-based metrics (1-NN, MMD, and Energy Distance), ensuring representativeness and generalization. The models achieved high performance, with out-of-sample coefficients of determination above 0.98. Interpretability was examined using explainable AI techniques based on SHAP values, indicating that temperature and equivalence ratio are key drivers of H₂ formation, whereas carbon and fixed carbon contents govern CO generation. The applicability domain was verified using Mahalanobis distance, ensuring prediction reliability. Finally, Differential Evolution optimization identified synergistic biomass–plastic pairs and operating conditions that maximize hydrogen production and syngas quality. The optimal solutions favored polypropylene-rich blends with lignocellulosic biomasses, yielding H₂ fractions of approximately 28%, H₂/CO ratios close to 1.1, and lower heating values around 6.5 MJ·Nm⁻³. The proposed approach integrates mechanistic modelling, explainable machine learning, and optimization, supporting the rational design of hydrogen-oriented co-gasification systems.