Integrated experimental, GA-ANN, and TOPSIS framework for optimizing mechanical and barrier properties of bio-modified PLA/wheat gluten film

Polylactic acid (PLA)/wheat gluten (PGW) films are a promising sustainable packaging material, but brittle. This work provides the first direct comparison between a synthetic modifier (acetic acid) and a natural bioactive modifier (rosemary essential oil, REO) within the same PLA/WG system, and also establishes a rigorously validated machine learning framework for small-dataset formulation optimisation by benchmarking five regression models using leave-one-out cross-validation (LOOCV), an approach well suited to the limited number of formulations (N = 6) typical of early-stage biomaterial screening. This validation showed that simple rest, gradient boosting, XGBoost) that tend to overfit, despite good in-sample perfomodels (linear regression, SVR) generalise much more reliably than ensemble methods (random formance. The best fitting model for tensile strength modified by REO was linear regression (R² = 0.978), whereas the best generalisation model for acetic acid modified properties was SVR (R² = 0.81–0.87). However, at this sample size, elongation at break and WVTR were not reliably predictable for either modifier. The multi-criteria ranking using TOPSIS revealed that 3% REO was the best formulation. Taken together, these results provide a mechanistic comparison of two modifier chemistries, and a validated approach to reliable ML-based formulation screening under realistic small-data conditions.