The rapid prediction of multi physical fields in heat sinks holds significant importance for chip cooling system fault diagnosis. Currently, the mainstream approach for multi physical fields prediction is computational fluid dynamics (CFD). However, the substantial time cost associated with traditional CFD calculations renders it impractical for online prediction. In this paper, a coupled multi-scale model based on CFD and convolutional neural networks (CNN) is proposed, which reduces the field prediction time to seconds. In this paper, a single-channel cold plate-based heat sink was studied. A comparison between the CFD-CNN model and traditional CFD was conducted and the results demonstrates that the maximum deviation of the temperature field error is 5%, with a predicted mean square error of only 0.77 K. For the pressure field, the maximum deviation is 15 Pa and the average error is less than 1 Pa. For the same case, the CFD-CNN calculation time is 0.5 s, which is 240 times faster than that of traditional CFD model.
 

Heat sink, CFD, CNN, Multi-Physical fields