Melt Temperature Estimation by Machine Learning Model Based on Energy Flow in Injection Molding.

Area of Science:

• Materials Science and Engineering
• Polymer Processing
• Artificial Intelligence in Manufacturing

Background:

• Accurate melt temperature measurement is critical for stable injection molding processes.
• Conventional thermocouples lack the responsiveness needed for real-time melt temperature monitoring.
• Existing methods struggle with the dynamic temperature changes during plasticizing.

Purpose of the Study:

• To develop machine learning (ML) models for predicting melt temperatures after plasticizing.
• To enhance prediction accuracy by incorporating various process and material parameters.
• To investigate the use of transfer learning for new material datasets.

Main Methods:

• Development of an optimized temperature sensor for reliable melt temperature data acquisition.
• Construction of three ML models: (1) using process settings, (2) adding material energy flow parameters, and (3) including specific heat and part weight.
• Evaluation of model performance using loss reduction and coefficient of determination; reliability analysis via SHAP values.

Main Results:

• The third ML model, incorporating specific heat and part weight, significantly outperformed others, with >70% loss reduction.
• The coefficient of determination improved by approximately 0.5 compared to the first model.
• A transfer learning model demonstrated high prediction performance and reliability with reduced dataset size for new materials.

Conclusions:

• ML models, particularly those integrating material-specific energy data, offer a highly reliable approach to predicting melt temperatures.
• The developed sensor and ML framework enable more stable and efficient injection molding.
• Transfer learning effectively reduces data requirements for adapting the models to new polymers.