Machine-learning prediction of biomass chemical composition using derivative thermogravimetric data of different...

The behaviour of biomass in bioenergy and biorefinery processes is determined by its contents of cellulose, hemicellulose, and lignin. The conventional wet chemical methods used to determine these fractions are reliable but time-consuming and laborious. This study attempts to determine whether these fractions can be adequately predicted from derivative thermogravimetric (DTG) data alone. A total of 75 biomass samples, including bamboo, agricultural residues, shells, and binary blends, were used. In addition, we calculated 13 simple, physically meaningful descriptors from each DTG curve (67 points from 27 to 687 °C), including peak height and temperature, areas under fixed-temperature windows, and area ratios. We trained seven algorithms, one for each component: Random Forest, Gradient Boosting, Extreme Gradient Boosting, Ridge, Partial Least Squares, Support Vector Regression, and k-Nearest Neighbours. All models were evaluated using nested leave-one-out cross-validation, with parameters optimised in the loop, and model stability was assessed with repeated fivefold cross-validation. The engineered descriptors improved every component. Cellulose was predicted with moderate accuracy (cross-validated R² of about 0.50 to 0.56, RMSE about 6.8%). Hemicellulose was weaker (R² about 0.38- 0.43, RMSE about 4.8%). Lignin could not be predicted reliably (stable R² about 0.13). The reason is physical: lignin decomposes slowly over a wide temperature range that overlaps with those of the other two components. SHAP and permutation analysis tied the predictions to sensible temperature regions. The study shows what DTG-based prediction can and cannot do across feedstocks, and it provides an open, reproducible pipeline.