Predictive Model Validation on a Small Experimental Biomass Dataset: A Comparative Study of Response Surface Methodology and XGBoost for Bio-Briquette Calorific Value

Abstract

Small experimental datasets remain a major limitation in biomass process modelling, where traditional empirical models may demonstrate excellent goodness-of-fit yet exhibit poor predictive reliability when applied to unseen conditions. This study evaluates the predictive modelling of the calorific value of bio-briquettes produced from hydrothermal liquefaction residue of breadfruit pulp using cassava starch as a binder, based on a previously published experimental dataset comprising three process variables (particle size, pressing pressure, and binder dosage) across seventeen experimental runs. The predictive performance of a classical Response Surface Methodology (RSM) model was compared with that of the Extreme Gradient Boosting (XGBoost) algorithm using Leave-One-Out Cross-Validation (LOOCV), a validation strategy particularly suited to small datasets. Although the quadratic RSM model achieved excellent agreement with the training data (R² = 0.932), its predictive performance declined substantially under LOOCV (R² = 0.097), indicating overfitting. Ridge regularisation partially improved model performance (LOOCV R² = 0.316), while XGBoost consistently achieved superior predictive generalisation (LOOCV R² = 0.569) with lower prediction errors (RMSE = 3.462 MJ kg⁻¹; MAE = 2.838 MJ kg⁻¹). Feature importance and correlation analyses identified binder dosage as the dominant factor influencing calorific value, whereas particle size and pressing pressure exhibited comparatively weaker effects within the investigated operating range. The integration of response surface visualisation with machine learning interpretation further provided complementary insight into factor interactions and predictive behaviour. The study demonstrates that rigorous predictive validation can reveal substantial differences between apparent model fit and true predictive performance in small experimental biomass datasets. Rather than replacing Response Surface Methodology, machine learning is shown to complement conventional experimental approaches by enhancing predictive reliability while preserving the interpretability of classical Design of Experiments. The proposed framework provides a practical approach for re-examining existing experimental datasets to support more reliable data-driven optimisation in biomass and bioenergy research.