Transfer learning based topology optimization of battery cooling channels design for improved thermal performance

In the field of battery thermal management, traditional cold plate design methods improve thermal performance through parameterization and structural optimization, but the high computational complexity results in a significant amount of time and computational resources required for the design of each cooling channel structure. This study proposes a method that combines transfer learning and Wasserstein Generative Adversarial Network (WGAN) gradient penalty (GP) to quickly generate novel battery cold plate structures with limited samples. Firstly, a small sample dataset of liquid cooled plate channels was constructed based on topology optimization, and the sample size was expanded through data augmentation. The WGAN-GP model is pre trained to learn cooling channel features, and then the generator parameters are transferred to formal training to generate new structures that traditional methods cannot obtain. Compared with traditional topology optimization, the training and generation time of WGAN-GP is only 6.38%. After verification through a three-dimensional electrochemical heat flux coupling model, the results showed that the cooling channel structure generated by WGAN-GP effectively reduced the average temperature (1.83 ℃), maximum temperature (2.41 ℃), and temperature standard deviation (0.382 ℃) of the battery, while reducing energy consumption by 7.91%, demonstrating the effectiveness of this method.