Process simulation and machine learning modeling of biomass wastes co-gasification for syngas and biochar production

Abstract

The urgency of climate change has accelerated research into renewable energy, including biomass co-gasification. While biomass waste offers a sustainable resource for producing syngas and biochar, the conversion process remains complex due to variability in feedstock properties, reactor design, and operating conditions. Additionally, traditional experimental and mechanistic modeling approaches are often time-consuming, costly, and limited in generalizability. This study addresses this gap by integrating experimental characterization, process simulation (PS), and machine learning (ML) to enhance understanding and prediction of autothermal biomass co-gasification outcomes. The primary objective of this research is to develop predictive models to optimize syngas and biochar production. Specifically, the study characterizes the physico-chemical and thermo-kinetic properties of five biomass feedstocks (coffee husks, groundnut shells, macadamia nutshells, rice husks, and tea wastes); develops a PS model to represent biomass co-gasification dynamics; develops and validates ML models to predict co- gasification outcomes; assess model robustness using new biomass blends; and evaluates the impact of model interpretability techniques on feature importance rankings. Proximate analysis method followed ASTM E1131-08, while ultimate analysis was conducted using a Carbon-Hydrogen-Nitrogen-Sulfur (CHNS) analyzer. Thermogravimetric analysis under a nitrogen atmosphere was employed to study thermal degradation, and kinetic parameters were estimated using the Coats-Redfern method. Aspen plus PS method was used to simulate a pilot-scale downdraft gasifier. ML models (Random Forest, Artificial Neural Networks, Gradient Boosting Regression, Support Vector Regression, and SuperLearner ensembles (SLE)) were developed and validated in MATLAB. Robustness was tested by validating the models with new feedstock blends. Interpretability was evaluated using permutation importance, Gini importance, and partial dependence plots. Proximate and ultimate analysis results revealed variability among feedstocks. Volatile matter (63.96% ±3.57%) indicated high syngas and tar potential, fixed carbon (19.62% ±2.69%) contributed to char formation, and carbon content (47.69% ±4.80%) suggested high energy conversion efficiency. Thermo-kinetic analysis showed peak devolatilization temperatures between 345°C and 380°C, activation energies ranging from 39 to 46 kJ/mol, and Gibbs free energy values between 151 and 162 kJ/mol, indicating favorable decomposition behavior. PS model achieved high accuracy with temperature deviations of 2°C (pyrolysis), 4°C (combustion), and 7°C (reduction), and syngas yield deviations of 0.21 Nm3/kg (Equivalence Ratio (ER) 0.17) and 0.34 Nm3/kg (ER 0.29). Sensitivity analysis showed that increasing ER enhanced hydrogen concentration by 10–15% and reduced carbon dioxide by 20–25%. All ML models performed well with Coefficient of Determination (R2) > 0.90 and Root Mean Square Error (RMSE) < 5%, confirming their effectiveness. Robust analysis showed that SLE maintained superior generalizability (R2 > 0.75, RMSE < 5%) for predicting char and syngas yields. However, hydrogen and cold gas efficiency predictions were less robust, indicating needs for more diverse datasets to improve generalization. Interpretability analysis identified ER and steam-to-biomass ratio as key predictors, with macadamia nutshells playing a critical role in enhancing char yield. In conclusion, this study demonstrates biomass waste co-gasification using integrated experiments, simulation, and ML, providing insights for sustainable energy. The findings recommend applying the developed ML models in academic research and practical gasification systems to support reliable process prediction and decision-making.