Graph Attention Network-Based Surrogate Model for Acceleration and Generalization of Injection Molding Simulations

Digital twin technology based on real-time simulation is required for advanced manufacturing in newer injection molding processes. However, the excessive computation time of conventional computer-aided engineering (CAE) simulations for injection molding limits real-time implementation. To address these limitations, this study developed a surrogate model to accelerate CAE simulations for injection molding with generalization performance for unseen geometric parameters and unseen injection molding process conditions. The graph attention network was used as the neural network. To ensure generalization performance for unseen geometric parameters, six types of geometric features representing the geometric characteristics of injection-molded products were defined and used for training. In addition, to ensure generalization performance for unseen injection molding process conditions, CAE simulation result data were generated by randomly sampling process conditions within a predefined range and used for training. The prediction was performed using the generated graph attention network-based surrogate model for models with unseen geometric parameters and injection molding process conditions that were not used in training. Results showed that the mean absolute percentage error values in the denormalized state were on average 5.2753%, 1.8758%, and 11.9311% for the fill time, and the temperature and pressure before the velocity-to-pressure switching point, respectively. In addition, compared with the existing commercial CAE simulation, the proposed surrogate model achieved approximately 44 times faster prediction for 50 different conditions of the injection molding process.