Use of machine learning methods for modelling mechanical parameters of PLA and PLA/native potato starch compound using aging data

Abstract Machine learning methods were applied to investigate changes in material properties during degradation, focusing on bio-based PLA and a PLA/native potato starch compound (50 wt%). Sixteen aging conditions involving various temperatures and humidity levels, and aging durations were examined. Characterization of aged samples involved tensile tests, FTIR analysis, weight or density measurements, and injection molding data. These data served as inputs to develop and compare predictive models of mechanical properties like Young’s modulus and elongation at break. Linear and polynomial regression, as well as multilayer perceptron (MLP) models were employed to evaluate their prediction accuracy. The best model accuracy (RMSE = 0.33) was achieved by segregating the dataset by material type and employing linear regression. Notably, employing two independent variables such as temperature and humidity led to high model quality (RMSE = 0.35). Effect diagrams revealed strong alignment between actual and modeled data, highlighting the comparative strengths of each modeling approach.