Applying machine learning to predict catalytic pyrolysis of palm biowaste fibers using TGA data

This study aims to predict the pyrolysis behavior of Chamaerops humilis fibers ( Ch Fs) and gain insights into their thermal degradation, using artificial neural networks (ANNs) as a data-driven modeling tool. Thermogravimetric analysis (TGA) was performed on Ch Fs between 30°C and 800°C at six heating rates (5, 10, 20, 30, 40, and 50 °C/min) under a nitrogen atmosphere. A feed-forward ANN with two input neurons (temperature and heating rate) and one output neuron (weight loss) was developed. Twenty-seven network architectures were tested to identify the optimal model, which was then used to calculate kinetic and thermodynamic parameters. Model-free methods (Flynn-Wall-Ozawa, Kissinger-Akahira-Sunose, and Starink) were employed for comparison. The best ANN model (ANN26, 5 × 17 × 1) achieved excellent prediction accuracy (R² > 0.99996) for all heating rates. The ANN successfully predicted weight loss trends, enthalpy (ΔH), and Gibbs free energy (ΔG), while slightly overestimating activation energy (Ea) in the KAS method. The results demonstrate strong agreement between ANN predictions and experimental data across all stages of pyrolysis. This work shows that even a simple ANN can accurately model the complex thermal degradation of Ch Fs, providing a reliable tool for process optimization. The approach enables prediction of pyrolysis behavior at different heating rates, highlighting the potential of ANNs for data-driven biomass conversion studies. • ANN26 predicted Ch Fs weight loss with R² > 0.99996 across all datasets. • TGA and DTA studied Ch Fs pyrolysis at β = 5–50 °C/min. • Hemicellulose, cellulose, and lignin decomposed at 220–500°C. • Kinetic parameters (Ea, ΔH, ΔS, ΔG) matched FWO, KAS, and Starink results. • ANN accurately predicted unseen data, aiding biomass pyrolysis optimization.