Sparse data-driven modelling of geometry adaptive heat transfer in multichip modules with model-agnostic meta-learning algorithm

Rapid thermal estimation for Multichip modules (MCMs) is crucial for structural design and optimization of electronic equipment. Although existing thermal estimation methods based on convolutional neural networks (CNN) can rapidly predict temperature distributions for various chip arrangements and power dissipations, they typically require large datasets and extensive training time, resulting in high computational and resource costs. To overcome these limitations, this paper proposes a sparse data-driven model integrating meta-learning with CNNs, specifically tailored for constructing geometry-adaptive heat transfer prediction models for MCMs, thus significantly reducing dependence on extensive training datasets. When adapting to new tasks, the proposed model requires only approximately 2 s of fine-tuning using merely 20 new data samples to achieve geometric adaptability, attaining an impressive prediction accuracy of approximately 99 %. This accuracy is comparable to conventional CNN-based surrogate models but reduces data requirements by approximately 80 %. Furthermore, the proposed model can estimate the temperature field within 10 ms, which is three to four orders of magnitude faster than traditional numerical simulations. The results demonstrate the model’s significant potential for efficient few-shot multitask learning in thermal estimation scenarios, substantially improving the utilization efficiency of historical MCM heat transfer datasets and effectively supporting real-time thermal estimation and rapid optimization of chip configurations.