Hybrid neural models for pressure control in injection molding

Industry standards place stringent requirements on the quality of injection-molded products. Existing models for quality prediction have a limited accuracy that is not adequate to deliver desired quality targets of three defects per million. Moreover, these quality control strategies often require large amounts of training data, while also becoming invalid with external process changes. This study compares the performance of three different models for prediction of the injection pressure: a conventional neural network, tuned on experimental data; a simulation network, trained on simulated data and fine-tuned on experimental data; and a hybrid network, which combines the training of neural networks with analytical knowledge of the molding process. The results indicate that the hybrid neural network performs especially well for a small number of training points. In particular, for training points ranging between one and four, the sum of squared errors (SSE) for a hybrid network is 13% to 33% of the SSE for a conventional network, and approximately 45% of the SSE for a simulation network. In addition, the hybrid network also shows the fastest convergence rate among the three models. However, as the number of training points increases, the conventional neural network outperforms all other models.