Machine learning driven multi-objective optimization of cellulose acetate film composition for food packaging applications.

Optimal proportions of plasticizers, crosslinkers, and hydrophobicity modifiers are essential for biopolymer film formulations. In this study, Cellulose acetate bioplastic films were prepared with varying concentrations of polyethylene glycol (PEG), malic acid (MA), and hexadecanoic acid (HAD). The resulting films were characterized for thickness (TH), water absorbency (WA), transparency (TP), and equilibrium moisture content (MC). Each of these attributes was modeled using artificial neural networks (ANN) and an ensemble regression tree. The ANN models offered an excellent fit to these responses (R 0.981-0.999, R: 0.928-0.978). The models were assessed using partial dependency, Local Interpretable Model-Agnostic Explanations, and Shapley explanation. Individual optimization projected a minimum TH of 0.03 mm, maximum TP of 43.14 % with WA, and MC below the detection limit. Multi-objective optimization suggested a blend containing 299.78 mg g PEG, 7.51 mg g MA, and 44.00 mg g of HAD while maintaining TH (0.06 mm), MC (0.05 %), and TP (42.83 %).