Transfer learning of machine learning models for multi-objective process optimization of a transferred mold to ensure efficient and robust injection molding of high surface quality parts

Surface quality is a main quality factor due to the importance of aesthetics, appearance, and perceived quality of products. Even when the process condition is optimized with the mold at the molding trial site, the process parameters still need to be adjusted again after the mold is transferred to the production site because of the dependency of surface quality on the molding machine, auxiliary equipment, and ambient conditions. In this study, transfer learning was employed to increase the efficiency of process optimization for high surface quality when the mold is transferred to a different molding site. Multi-task artificial neural networks (ANN) for surface gloss and defect prediction were trained by the dataset from the original production site. The pre-trained ANN model was then transferred together with the mold to a different production site. The pre-trained model gave acceptable prediction performance of R² = 0.94 on surface gloss for the new machine but performed poorly for surface defect prediction due to different machine characteristics. The transfer learning on the single trainable output layer exhibited a high and stable prediction performance. Application of transfer learning not only delivered a better surface gloss prediction (R² > 0.95) and a similar prediction on surface defect (accuracy 0.90) than a typical machine learning approach without transfer learning, but also reduced the required dataset size by about 50 %. The transferred model enabled multi-objective, model-based optimization of process parameters that led to robust and efficient production of high surface quality injection molded parts, which was verified by physical molding experiments.