Image-based assessment and machine learning-enabled prediction of printability of polysaccharides-based food ink for 3D printing

Abstract

Despite the growing demand and interest in 3D printing for food manufacturing, predicting printability of food-grade materials based on biopolymer composition and rheological properties is a significant challenge. This study developed two image-based printability assessment metrics: printed filaments’ width and roughness and used these metrics to evaluate the printability of hydrogel-based food inks using response surface methodology (RSM) with regression analysis and machine learning. Rheological and compositional properties of food grade inks formulated using low-methoxyl pectin (LMP) and cellulose nanocrystals (CNC) with different ionic crosslinking densities were used as predictors of printability. RSM and linear regression showed good predictability of rheological properties based on formulation parameters but could not predict the printability metrics. For a machine learning based prediction model, the printability metrics were binarized with pre-specified thresholds and random forest classifiers were trained to predict the filament width and roughness labels, as well as the overall printability of the inks using formulation and rheological parameters. Without including formulation parameters, the models trained on rheological measurements alone were able to achieve high prediction accuracy: 82% for the width and roughness labels and 88% for the overall printability label, demonstrating the potential to predict printability of the polysaccharide inks developed in this study and to possibly generalize the models to food inks with different compositions.