Data from: Physically interpretable surrogate modeling of thermal fields in electronics cooling using combined proper orthogonal decomposition and neural networks

As advances in semiconductor technology drive higher power densities,
thermal management is becoming increasingly crucial to maintaining the
performance and reliability of devices. In recent decades, computational
fluid dynamics (CFD) has been widely used for thermal management system
design. However, the computational requirement of CFD limits its use for
comprehensive system optimization and digital twin active control systems,
which require quick, high-fidelity thermal field prediction. Data-driven
surrogate models address the speed limitations of CFD but are often
treated as black boxes whose predictions are physically uninterpretable.
Neural networks (NN) combined with proper orthogonal decomposition (POD)
maintain the speedup of other surrogate models while offering a physically
interpretable latent space. The interpretability of POD-NN learned
representations remains largely unexplored. This work develops and
demonstrates a POD-NN surrogate model framework to predict 2D thermal
fields for a liquid-cooled dual-chip cold plate with four retained POD
modes. The resultant model predicts fields 412,500 faster than
CFD. An error decomposition analysis is used to assess the contribution of
POD-based and NN-based error across a range of dataset sizes and identify
an error floor dependent on the number of modes retained rather than
dataset size. A systematic Pearson correlation analysis is used to link
all four POD coefficients to identifiable physical features, and thermal
resistance is mapped in latent space. The presented framework enables
sensitivity analysis, design optimization, and active thermal control
within the physics-aware POD coefficient space, and advances POD-NN
architectures from accelerators to interpretable engineering design tools.