Work

A mask-aware Transformer framework for cross-topological thermal prediction in 2.5D chiplet systems

Year: 2026
Type: article
Abstract: 2.5D chiplet heterogeneous integration systems are key for extending Moore's Law and boosting integrated circuit performance. They are widely used in fields such as artificial intelligence training and high-performance computing. However, their dense multi-chip structure creates new and significant thermal management challenges. Current deep learning models based on fixed-grid architectures, such as U-Net, have limited local receptive fields and struggle to explicitly model global, long-range thermal coupling effects across variable topological layouts. To address this, this paper proposes HeatTransUNet, a deep learning framework based on a mask-aware Transformer. The framework combines variable-length sequence modelling with spatial physical field generation, transforming physical heat sources into flexible, discrete-sequence entities. By leveraging global self-attention, the model learns global dependencies among chiplet tokens at the sequence level. A dynamic masking strategy is employed to mathematically isolate irrelevant padding noise. Subsequently, a U-Net spatial decoder utilizes feature smoothing to reconstruct these sequence features, producing high-precision, continuous 2D steady-state temperature fields. HeatTransUNet achieves an average absolute error of 0.028 K on a benchmark dataset of 6000 cases with six distinct topologies, corresponding to only 0.05% of the temperature span across the complete dataset. In few-shot transfer to four target layouts excluded from source-domain pretraining, HeatTransUNet was fine-tuned using 320 pooled target-domain training samples drawn from 100 generated cases per layout, reducing prediction error by approximately 50% compared with training from scratch across five independent target-domain data splits. This study provides a highly adaptable, high-precision, and low-cost electro-thermal co-design solution for modern complex electronic devices during early agile design.
Language: English
Open Access status: gold
Estimated APC paid: $1,620
OpenAlex ID: W7237338191
FWCI:
Cites: 39
Cited by:
Related to: