| dc.description.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 |
en_US |