Date of Award
5-2026
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Automotive Engineering
Committee Chair/Advisor
Saeed Farahani
Committee Member
Laine Mears
Committee Member
Ramy Harik
Committee Member
Cheoljoon Jeong
Abstract
Plastics and composites manufacturing encompasses a broad family of processes, including injection molding, resin transfer molding, compression molding, and extrusion, that together account for the vast majority of polymer and composite component production worldwide. These processes are inherently challenging due to the multi-physics nature of polymer processing, the vast possible combinations of resins, reinforcements, and additives, and the strong dependence of final part properties on process history. Among these processes, injection molding is one of the most widely deployed, producing high-volume components across automotive, medical, consumer electronics, and packaging industries. However, accurately monitoring and predicting process behavior in injection molding remains challenging due to the complex, multi-physics, multi-scale nature of this process and the limited visibility of real-time phenomena happening within the mold cavity. Conventional modeling approaches (physics-based simulations and data-driven models) exhibit complementary strengths but suffer from inherent limitations when applied independently. Physics-based models provide interpretability and design-stage insight but rely on ideal assumptions and static inputs, while data-driven models capture nonlinear process behavior but lack generalizability and consistency. This dissertation investigates hybrid modeling frameworks that integrate physics-based and data-driven methods to address these persistent challenges. The central hypothesis is that the strategic integration of these complementary modeling paradigms enables improved process monitoring, enhanced simulation accuracy, and more reliable energy predictions than standalone approaches. Three hybrid modeling frameworks are developed and experimentally validated, each targeting a distinct manufacturing challenge in injection molding. A cross-method performance study first establishes the quantitative justification for hybrid modeling by comparing the design of experiments combined with regression analysis, physics- based simulation, and machine learning across an experimental dataset. This study demonstrates the strengths and shortcomings of each molding approach and highlights the distinct advantages that allow each method to complement the other, emphasizing the importance of developing hybrid modeling frameworks for plastic/composites manufacturing applications. The first framework develops a physics-constrained long short-term memory (PC-LSTM) soft sensor that infers cavity pressure from exterior surface-mounted strain measurements on legacy molds. By embedding physics-based knowledge from finite element analysis into the training process through a composite loss function, the PC-LSTM achieves cross-material generalization from polypropylene (PP) to polymethyl methacrylate (PMMA) without requiring experimental data from the target material. The second framework introduces Manufacturing Asset Profiles (MAPs), a structured methodology for characterizing machine, mold, and material behavior from sensor data and translating these profiles into refined simulation input parameters for Moldex3D, systematically reducing discrepancies between simulation predictions and experimental measurements. The third framework combines a Hidden Markov Model (HMM) with a rule-based stochastic decision-making algorithm to dynamically select between physics-based and data-driven energy predictions based on the inferred operating state, achieving 85.51% accuracy across five material types, including virgin and recycled polymers. These contributions establish hybrid modeling as a generalizable and scalable paradigm for advancing process understanding and control in plastics and composites manufacturing. The results demonstrate that hybrid frameworks can enhance process monitoring, improve simulation accuracy, and enable adaptive prediction under real-world variability. The frameworks are designed for deployment on edge computing devices and are transferable to broader manufacturing vi domains. This work provides a foundation for advancing and supporting the development of more efficient, adaptive, and sustainable plastics and composites manufacturing systems.
Recommended Citation
Shannon, Luke, "Modernizing Plastics and Composites Manufacturing using Hybrid Modeling Techniques" (2026). All Dissertations. 4268.
https://open.clemson.edu/all_dissertations/4268