Small Sample Machine Learning for Predicting Composition Changes in Chemically Pretreated Kapok Fiber

Abstract

Kapok is a lightweight hollow tropical fiber with potential as a renewable feedstock, but its waxy surface and lignin-containing matrix complicate aqueous processing. This study evaluated whether limited-data machine learning could predict kapok fiber composition after aqueous chemical pretreatment using only three process descriptors. Kapok fibers were pretreated using dilute H2SO4, hydrothermal water, and NaOH across a shared window of residence time, temperature and pH-associated pretreatment environment. A data set of 33 pretreatment runs was constructed, with cellulose, hemicellulose and lignin contents used as composition targets after correction with solid yield. Single-output and wrapper-based multitarget regression models were screened and then evaluated using leave-one-out cross-validation to reduce dependence on a single train–test split. The LOOCV results showed target-dependent predictability. In single-output modeling, cellulose was the most predictable component, with GBDT giving the strongest performance (R2 = 0.782; RMSE = 4.185 percentage points), whereas hemicellulose and lignin showed weaker prediction. In wrapper-based multitarget modeling, GBDT improved simultaneous prediction of cellulose, hemicellulose and lignin, giving R2 values of 0.909, 0.653, and 0.672, respectively. Feature-importance analysis indicated that temperature was the dominant descriptor for polysaccharide-related composition changes, whereas the pH-associated pretreatment environment contributed more strongly to lignin-related variation. These findings position limited-data machine learning as a useful exploratory tool for modeling pretreatment–composition relationships in kapok fiber, while also highlighting the need for cautious validation when sample size is small.