Machine learning-driven modeling framework for steam co-gasification applications

Steam co-gasification of biomass and plastic waste is a promising route for syngas production and waste valorization. However, accurately predicting syngas composition remains challenging due to inherent complexity and nonlinearity of the process. This study presents a comprehensive comparative analysis between conventional process simulators-based models (Aspen Plus), namely the thermodynamic equilibrium (TEM), restricted thermodynamic (RTM), and kinetic (KM) modeling approaches, and machine learning (ML) models for the prediction of the syngas composition. Using 208 experimental data points compiled from 20 published studies covering various feedstocks and gasification conditions in bubbling fluidized bed gasifiers (BFBG), the performance of the models was evaluated after extensive data preprocessing. Among several ML algorithms evaluated, the neural network (NN) delivered the lowest average root mean square error in syngas mol fraction predictions (0.0174), outperforming RTM (0.0966), KM (0.1378), and TEM (0.1470). To explore input–output relationships beyond interpolation, a conditional generative adversarial network (cGAN) generated synthetic data, which served as the basis for sensitivity and interpretability analyses. The NN, acting as a surrogate model, was paired with SHapley Additive exPlanations (SHAP) and Partial Dependence Plots (PDP) to quantify the effects and nonlinear interactions of key features on syngas yields providing actionable insights for process optimization. • Experimental studies compiled and dataset curated for model development. • Screening of multiple ML models for top-performing surrogate model. • Comparative assessment conducted between TEM, RTM, KM, and ML models. • Synthetic data generated using cGAN to support extended sensitivity analysis. • SHAP and PDP analyses employed to determine key process variables.