Comparing pretraining methods with different fidelities for a 2D cooling problem in injection molding

Abstract Surrogate models leveraging Neural Networks (NNs) have demonstrated significant potential in accurately modeling the intricate dynamics of the injection molding process, enabling the prediction and optimization of process and material parameters on internal thermodynamics and final product quality. However, the substantial effort required for data collection remains a notable challenge. This study investigates the effectiveness of pretraining methods with varying fidelities for a two-dimensional cooling problem in injection molding. Specifically, NNs are pretrained using low-fidelity (LF) solutions derived from the Finite Element Method (FEM) and Physics-Informed Neural Networks (PINNs). These pretrained models are subsequently fine-tuned with high-fidelity (HF) simulation data to enhance prediction accuracy across previously untrained scenarios. While both FEM and PINNs provide approximations of the cooling process, PINNs offer a continuous solution space, however, at a higher computational cost compared to FEM. Respectively, FEM is more computationally efficient but less proficient at handling highly parameterized input spaces. Comparing different fidelities reveals that prioritizing the capture of main effects over detailed physical implementation during pretraining can be advantageous. The fine-tuned surrogate model demonstrates robust approximation capabilities with limited simulation runs, highlighting its data efficiency.