A study on a rapid prediction method for transient temperature fields in electronic devices based on block-wise POD and nonlinear mapping

Real-time transient thermal simulation of high-power-density electronic devices involves a fundamental trilemma among local hotspot resolution, computational latency, and mathematical transparency. This paper proposes a non-intrusive reduced-order model (NI-ROM) that integrates block-wise proper orthogonal decomposition (Block-wise POD) with physically motivated polynomial feature mapping. The block-wise strategy partitions the computational domain by material properties, enabling independent mode extraction within each subdomain and eliminating inter-media mode mixing. In the latent space, a second-order polynomial regression is used to match the nonlinear dynamic temperature changes of transient heat conduction. The model is validated through high-fidelity CFD simulations and a custom-built multi-condition experimental test bench under ambient temperatures from 30 °C to 50 °C. Results demonstrate that the prediction error at critical heat-generating components is consistently below 4.5%, while a complete three-dimensional full-field reconstruction for a target-state query is completed within approximately 9.31 s, achieving a computational speedup of roughly 710× over the corresponding transient CFD simulation. The proposed framework thus offers a transparent and computationally efficient paradigm for transient thermal analysis, and is potentially applicable to online thermal management of electronic systems under specific structural configurations.