Machine-Learning-Guided Experimental Prioritization of Seaweed-Derived Bioplastic Films Toward an LDPE-like Mechanical Target Using Sparse Literature Data

Abstract

Seaweed polysaccharide films are potential alternatives to petroleum-derived flexible packaging, but literature data are sparse and heterogeneous. This study evaluates whether composition-only machine-learning models can prioritize reported seaweed formulations for experimental follow-up near a nominal low-density polyethylene (LDPE) mechanical target of 15 MPa tensile strength and 300% elongation at break. A published dataset of 115 formulations with 41 compositional predictors was analyzed using a regularized linear baseline (ridge regression), random forest, and a deliberately shallow, regularized XGBoost model. Model performance was estimated using repeated five-fold cross-validation (10 repeats; 50 test-fold evaluations). For tensile strength, mean R2 was 0.617 ± 0.174 for random forest and 0.628 ± 0.150 for XGBoost, with corresponding RMSE values of 12.35 ± 3.52 and 12.27 ± 3.84 MPa. For elongation, random forest and XGBoost reached mean R2 values of 0.498 ± 0.194 and 0.470 ± 0.188, respectively. The linear baseline was materially less stable, indicating that the available composition-property relations are not adequately represented by a global linear model. The maximum observed elongation was 172.5%, which is 42.5% below the 300% screening target; equivalently, the target lies 73.9% above the dataset maximum. Consequently, the model cannot establish LDPE equivalence or credible extrapolation to the target. Residual diagnostics, learning curves, and SHAP analyses were added to bound interpretation. The defensible outcome is experimental prioritization within the observed domain, not inverse design or material substitution.