Predictive modeling of thermogravimetric analysis data of medium-density fibreboard waste: a machine learning approach

Abstract

This study investigates the application of machine learning (ML) techniques to predict thermogravimetric analysis (TGA) mass-loss data of medium-density fibreboard (MDF) waste under oxidative and pyrolytic conditions. Thermal decomposition provides a potential alternative to landfilling by reducing harmful additives and redistributing nitrogen-containing species among char, condensable products, and gases. Experimental data were obtained at four heating rates (5, 10, 15, and 20 °C/min) from room temperature to 600–800 °C. Five regression models were evaluated, with Linear Regression (LinReg) and Partial Least Squares Regression (PLSR) serving as baseline linear models, and k-Nearest Neighbors (k-NN), Random Forest (RF), and Gradient Boosting Regression (GBR) applied as the main nonlinear models. The key contribution of this work is the comparison of five regression models for MDF-waste TGA prediction under oxidative and pyrolytic environments across 3 different scenario-based data splits. Three modeling scenarios were evaluated using graphical tuning and Bayesian optimization. Scenario 1, based on a randomized 80/20 split of the complete dataset, is interpreted as an interpolation baseline. Scenario 2, which involved training on 5 and 10 °C/min and testing on 15 and 20 °C/min, provided the most credible heating-rate-transfer configuration, with the selected GBR models achieving test R² values of 0.989 for oxidation and 0.997 for pyrolysis. Scenario 3 which involved training on 5 °C/min and testing on 10, 15 and 20 °C/min, showed weaker curve-level alignment, particularly in the char-oxidation and terminal residual-mass regions. These results indicate that tree-based models, supported by Bayesian optimization, can provide an effective computational approach for MDF thermal-degradation prediction under the studied conditions, although model applicability remains material and condition-specific and is limited to the investigated MDF waste, atmospheres, heating rates, and experimental conditions.