DiffCool: Label-Free Synthesis of Chip-Tailored Heat Sinks via Thermal-Aware Diffusion

Published in The 40th Annual Conference on Neural Information Processing Systems (NeurIPS), 2026

Recommended citation: S.Y. Liang, Z.X. Wang, C.H. Wang, S.Y. Li, Y.S. Zhang, L.L. Jin, Z. Zhuang, U. Schlichtmann, B. Yu, T.-Y. Ho, "DiffCool: Label-Free Synthesis of Chip-Tailored Heat Sinks via Thermal-Aware Diffusion," The 40th Annual Conference on Neural Information Processing Systems (NeurIPS), 2026.

Thermal management is a key concern in chip and package design, but heat sinks are often chosen from standard form factors rather than designed for the thermal behavior of a specific chip. Tailoring a heat sink to an individual chip is difficult, in part because the designs needed for supervised learning are generally not available.

This paper presents DiffCool, a label-free synthesis approach that generates chip-tailored heat sinks using diffusion models guided by thermal information. Because it does not rely on labeled design pairs, the approach targets settings where ground-truth heat sink designs are not available, while aiming to align the generated geometry with the thermal behavior of the target chip.