Abstract
This paper proposes a combination of Proper Or thogonal Decomposition (POD) and Conditional Generative Ad versarial Neural Networks (CGaNet) to predict the steady-state temperature distribution of PCB and optimize heatsink design under forced convection conditions. In comparison to supervised models, the POD-CGaNet offers the advantages of accurate predictions over a wide data range while relying on a small range of training data. In contrast to an unsupervised temper ature prediction model, the POD-CGaNet features a few model parameters, low hardware requirements, and short training time. Furthermore, to explore how input dimensionality within the extrapolation range influences prediction accuracy, this work demonstrates that POD-CGaNet can achieve an accuracy of R2 = 0.9 even when all input parameters fall within extrapolated ranges, outperforming the other models in comparison. The prediction results are consistent with numerical analysis and ex perimental results under two different cases. Due to its low error and fast prediction, POD-CGaNet can be integrated into the Non dominated Sorting Genetic Algorithm II (NSGA-II) framework as a surrogate model instead of commercial software in heatsink design process. Within large design space, the algorithm can reduce 95% optimization time compared to FEM methods.
Funding Information
  • European Union’s Horizon Europe (101056857)