Thermal management has become a critical challenge in advanced semiconductor packaging, where high power density and complex three‐dimensional integration hinder direct temperature measurement and increase the risk of performance degradation. This study presents a scalable and power‐efficient framework for real‐time, die‐level temperature prediction by integrating finite element simulations, machine learning (ML), and hardware acceleration on a reconfigurable processor. A compact thermal model was first calibrated using limited infrared thermography data through hybrid ML–Pareto optimization, ensuring balanced accuracy across multiple power conditions. The optimized simulations generated high‐fidelity synthetic datasets, which—combined with empirical measurements—were used to train a lightweight single‐layer perceptron capable of predicting heat maps from surface‐temperature inputs. Implemented on a Xilinx Zynq platform, the model achieved inference latencies comparable to a desktop GPU while consuming only 1.71 W—delivering more than 21 times greater energy efficiency than the GPU and over 2600 times faster computation than full finite element simulations. The proposed approach enables accurate, real‐time thermal monitoring in resource‐constrained environments, with strong potential for deployment in edge artificial intelligence, mobile systems, and Internet of Things devices. This framework also opens pathways for integrating predictive thermal control into next‐generation intelligent electronic systems.
Lee et al. (Sun,) studied this question.