Multi-Task Gaussian Process Prediction of Pretreatment-Induced Compositional Changes in Kenaf Core Under Data-Limited Conditions

Pretreatment-induced compositional changes in lignocellulosic biomass are difficult to predict when experimental datasets are small and heterogeneous. This study applied multi-task Gaussian process regression (MTGPR) to jointly predict cellulose, hemicellulose, and lignin contents from residence time, temperature, and initial pH for 31 kenaf-core samples across autohydrolysis, dilute-acid, and alkaline pretreatment. MTGPR attained the lowest mean per-component RMSE under leave-one-out cross-validation (2.54 percentage points), 18% below independently fitted single-task GPRs (3.09 percentage points) and 8% below a response-surface model (2.77 percentage points), while matching the strongest single-task learner, random forest (2.61 percentage points). A diagonal-B ablation yielded 2.51 percentage points, showing that this gain stems from joint fitting with a shared kernel rather than from the explicit cross-task covariance term. At support = 8, a simple route-mean baseline outperformed all learned models, and under leave-one-route-out validation no learned model surpassed the route-mean reference. Cross-conformal intervals restored coverage to 99.5% but at ±7.38 percentage points half-width, which is over-conservative for decision-making. Joint fitting is therefore beneficial relative to independent fitting under data scarcity, but the benefit derives from shared structure rather than cross-task correlation, it does not extend to unseen pretreatment routes, and the uncertainty estimates still require recalibration.