Quick Prediction of Complex Temperature Fields Using Conditional Generative Adversarial Networks

Full Text in Publisher | Download Paper

• This study addresses the computationally intensive challenge of thermal management design in electronic devices by introducing an end-to-end physical field prediction approach based on conditional generative adversarial networks. The model establishes a mapping between geometric structures and temperature fields, enabling accurate prediction of peak temperatures and visual generation of detailed physical field distributions. The research further investigates the impact of training data volume on model performance and demonstrates that the approach exhibits strong generalization through fine-tuning, even when training data are limited or applied to novel devices. Its core advantage lies in extremely high computational efficiency, achieving orders-of-magnitude acceleration over traditional grid-based simulations while significantly reducing energy consumption. The flexible geometric encoding system provides an efficient solution for thermal analysis and optimization across a wide range of electronic devices.