Analysis of Geometric Feature Importance for Enhancing Performance of Injection-Moulding Simulation Surrogate Model

Abstract Injection-moulding computer-aided engineering (CAE) simulation is a numerical analysis process with high computational costs, presenting limitations for direct application to real-time simulation-based digital twin technology. To overcome these limitations, this study proposes an accelerated surrogate model that ensures generalisation performance for unseen geometries and process conditions within the same product family, and conducts a feature importance analysis for model interpretability. The model was trained using simulation data generated by randomly sampled combinations of six injection-moulding process conditions across various toothbrush geometries. In addition, seven geometric features based on distance, area, volume, thickness, and curvature were utilised as input features to effectively represent the characteristics of the product geometry. The surrogate model was constructed using a graph attention network that integrates an attention mechanism with a message-passing structure. The proposed model was applied to predict unseen product geometries and process conditions that were not included in the training dataset. The model achieved normalised root mean square errors (NRMSE) of 0.0125, 0.0451, 0.0712, and 0.0373 for fill time, the temperature at the end of filling, the pressure at the velocity-to-pressure switchover point, and the average of the three directional deflection components, respectively, averaged over 400 unseen process-condition combinations. In terms of the average acceleration performance across the four toothbrush geometries, the unseen product geometry preprocessing accounted for approximately 20.68% of the computation time of the conventional simulation, whereas prediction time for unseen process conditions required on average only approximately 0.14%. In addition, the importance of the geometric features for the prediction accuracy of each physical quantity was analysed using permutation feature importance. Based on this analysis, key geometric features were identified, and an ablation study was conducted to further examine their influence on the prediction of injection-moulding physical quantities. These results provide useful insights into geometric feature selection for injection-moulding surrogate models.