Therm-FM: Foundation Model is ALL YOU NEED for 3D-ICs Thermal Simulation

2026-05-21

AI Summary

This work addresses the limited reusability of traditional data-driven thermal simulation methods for 3D integrated circuits (3D-ICs), which rely heavily on large volumes of high-fidelity finite element data. To overcome this challenge, the authors propose Therm-FM, the first approach to transfer a pre-trained partial differential equation (PDE) foundation model to 3D-IC thermal simulation. By integrating neural operators with multi-fidelity learning and introducing a thermal equivalent modeling mechanism, Therm-FM enables highly accurate and efficient prediction of both steady-state and transient thermal fields using only minimal target-domain data. Evaluated on public and industrial-scale benchmarks, Therm-FM surpasses state-of-the-art methods with less than 20% of the training data, reducing average errors by up to 10.6×. Moreover, cross-chip adaptation requires merely 10–30 samples to match or exceed the performance of full-data baselines.

Abstract

Data-driven thermal predictors for 3D-ICs are often trained from scratch for each chip design using many high-fidelity finite-element simulations, leading to high data-generation cost and costly cross-design reuse. We propose Therm-FM, a neural operator framework that adapts a pretrained partial differential equation (PDE) foundation model to steady-state and transient 3D-IC thermal simulation. The motivation is that steady-state and transient chip-level heat conduction respectively share elliptic and parabolic operator structures with diffusion-type PDEs, allowing pretrained diffusion priors to provide an effective initialization for thermal-field prediction under heterogeneous materials, dense TSV/microbump interconnects, and package-level boundary conditions. To further reduce data-generation cost, Therm-FM incorporates a thermal-equivalent multi-fidelity training strategy that uses low-cost approximate simulations for thermal-domain adaptation and limited high-fidelity samples for calibration. Experiments on public HotSpot benchmarks and industrial 3D-IC package benchmarks show that Therm-FM achieves up to a 10.6x reduction in mean error and surpasses prior best accuracy with less than 20% of the training data. In cross-chip adaptation, it matches or surpasses full-data baselines in several metrics using only 10--30 target samples. We release datasets, source code, and pretrained models at https://github.com/haiyangxin/Therm-FM.

Problem

• 3D-ICs
• thermal simulation
• data-driven modeling
• high-fidelity simulation
• cross-design reuse

Innovation

• foundation model
• neural operator
• 3D-IC thermal simulation
• multi-fidelity learning
• PDE adaptation