Comparison of Hybrid Machine Learning Approaches for Surrogate Modeling Part Shrinkage in Injection Molding.

Area of Science:

• Materials Science and Engineering
• Manufacturing Processes
• Computational Science

Background:

• Machine learning (ML) offers potential for modeling complex, non-linear injection molding processes.
• Challenges in data collection and the dynamic nature of injection molding hinder purely data-driven ML applications.
• Integrating process knowledge with data is crucial for robust predictive models.

Purpose of the Study:

• To compare hybrid modeling approaches that combine process data with physical knowledge for injection molding.
• To evaluate the effectiveness of various hybrid strategies, including feature learning, fine-tuning, and physical constraints.
• To identify the optimal hybrid approach for predicting part quality using both simulated and experimental data.

Main Methods:

• Developed and compared hybrid machine learning models integrating process data with constitutive equations and numerical simulations.
• Investigated approaches such as feature learning, fine-tuning, delta-modeling, preprocessing, and physical constraints.
• Trained and validated models using experimental and simulated shrinkage data from an injection-molded part.

Main Results:

• All hybrid approaches demonstrated superior performance compared to purely data-based models.
• The fine-tuning approach achieved the best predictive accuracy in the simulation setting.
• A combination of feature learning (physical model calibration) and physical constraints yielded the best results in the experimental setting.

Conclusions:

• Hybrid modeling strategies significantly enhance the prediction of injection molding process behavior and part quality.
• The optimal hybrid approach depends on the data context, with fine-tuning excelling in simulations and a combined approach favored for experimental data.
• Integrating physical knowledge into ML models is essential for overcoming data limitations and improving predictive accuracy in manufacturing.