Machine learning-guided tuning of shape memory behavior in 4D-printed single-layer elastomer strip

Purpose This study introduces a novel four-dimensional printing (4DP) methodology employing biocompatible blends of polyurethane (PU), Eudragit® S100 and Metronidazole (MTZ) integrated with machine learning (ML) algorithms capable of modeling intricate and nonlinear relationships to predict critical response variables. Design/methodology/approach This study focused on predicting the ultimate tensile strength (UTS), shape memory programming, and activation stages using a dataset of 72 instances per response. Key printing parameters, including nozzle temperature, raster angle and printing speed, were optimized to enhance both mechanical properties and shape memory performance. The initial physical mixtures of PU, Eudragit® S100 and MTZ were processed into shape-memory filaments via hot-melt extrusion. Linear regression of multiple ML models, least absolute shrinkage and selection operator, ridge, elastic net, k-nearest neighbors (k-NN), decision tree (DT), random forest (RF), extra trees, support vector regression (SVR), response surface methodology, stacking, voting ensembles and artificial neural networks (ANN) were evaluated for predictive accuracy. Findings The pH-responsive nature of the Eudragit® S100 matrix was exploited to highlight its pharmaceutical potential, particularly for site-specific drug delivery. The experimental results revealed a significant shape-recovery ratio of 89.43% and a UTS of 11.15 MPa. A three-dimensional printed capsule fabricated under optimized conditions demonstrated excellent mechanical stability and shape memory behavior, indicating its future suitability for targeted drug delivery in the gastrointestinal tract. Originality/value This study emphasizes the integration of ML with 4D-printed PU, Eudragit® S100 and MTZ composites; the experimental outcomes of UTS and shape recovery and pharmaceutical relevance for futuristic colonic drug delivery.