Enhancing Injection Molding Optimization for SFRPs Through Multi-Fidelity Data-Driven Approaches Incorporating Prior Information in Limited Data Environments

Injection molding is one of the dominant methods for mass-producing short fiber reinforced plastics renowned for their exceptional specific properties. In the utilization of such composite components, optimization of process parameters significantly influences material characteristics and part performance. However, in industrial practice, this process often relies on intuition and iterative experimentation. Prior studies have introduced data-efficient optimization methods but faced limitations in adopting minor variations in the product development cycle. This study introduces a multi-fidelity optimization framework aimed at efficiently addressing new problems by leveraging previously acquired knowledge from analogous domains, particularly accommodating alterations in material scenarios. Two data-driven frameworks are explored: 1) Gaussian process-based and 2) neural network-based, each employing distinct information-transferring techniques, hierarchical Kriging and transfer learning, respectively. Bayesian optimization of process parameters under limited data budget, which is typical in realistic industrial settings, is performed. The results highlight the efficiency of the proposed framework, demonstrating superior performance not only in data-driven modeling but also in optimization efficiency compared to conventional single-fidelity approaches. The Pearson correlation coefficient is utilized to assess the applicability of the multi-fidelity framework in handling the inherent ambiguity of the similarity of problem scenarios. The proposed method is believed to be adaptable and versatile, offering potential application across various challenges in process optimization. This research introduces a multi-fidelity (MF) data-driven modeling and optimization approach for injection molding process optimization, leveraging prior data accumulated during product development cycle. Emphasizing the realistic industrial challenge of limited data availability, the data efficiency of the MF frameworks, specifically hierarchical Kriging and transfer learning, are systematically evaluated. The proposed framework demonstrates elevated performance and data-efficiency compared to conventional single-fidelity approaches. image