Development of Digital Twins to Increase the Efficiency and Sustainability of the Injection Molding Process

The injection molding industry faces a persistent efficiency ceiling during process setup, where parameter tuning remains a knowledge-intensive, trial-and-error procedure reliant on the tacit expertise of operators. This bottleneck results in excessive scrap, extended downtimes, and variability in product quality. Conventional Lean methodologies, such as Single-Minute Exchange of Die (SMED), address the procedural aspects of changeovers but fail to capture and operationalize the cognitive, corrective logic of expert decision-making. At the same time, recent advances in Industry 4.0—Digital Twins, physics-informed machine learning, and AI-driven knowledge capture—offer promising tools, though their application to this specific challenge remains fragmented. This thesis proposes, develops, and validates a novel Digital Twin framework that integrates physics-informed machine learning, sim-to-real transfer learning, and human-in-the-loop knowledge acquisition to overcome the setup efficiency ceiling in injection molding. The research follows a structured, multi-phase methodology. First, foundational studies demonstrate that predictive models of process physics can be trained with drastically reduced experimental burden by leveraging physics-informed neural networks and transfer learning from simulations. Second, a robust automated inspection pipeline is developed, employing self-supervised vision models and transformer-based architectures to achieve accurate defect detection under data-scarce conditions. Third, a knowledge acquisition workflow is introduced, systematically coupling process parameters (State) with visual defect outcomes (Outcome) to codify expert corrective logic. These components are integrated into a cognitive Digital Twin capable of both predictive quality estimation and prescriptive parameter optimization. The framework is validated through a comprehensive industrial case study. Five experimental trials of the Digital Twin-assisted setup process are benchmarked against a historical baseline of expert-driven sessions. Results show statistically significant improvements, with reductions in both total setup time and scrap produced. Beyond these operational KPIs, surface profilometry analysis using a Sensofar S neox confocal profiler confirms the physical quality of molded parts, demonstrating that weld lines in assisted trials exhibit minimal visibility and excellent surface topography. Together, these results provide robust empirical evidence that the Digital Twin not only accelerates process stabilization but also ensures defect-free outcomes. The contributions of this work are threefold: (1) a data-efficient methodology for predictive modeling in data-scarce manufacturing contexts, (2) a validated architecture for codifying tacit expert knowledge into a transferable Digital Twin, and (3) the first empirical demonstration of a “SMED 4.0” solution that bridges the gap between lean operational theory and advanced Industry 4.0 technologies. By addressing both the scientific and practical challenges of parameter-setting, this research provides a replicable framework for enhancing efficiency, sustainability, and resilience in high-variety manufacturing.