Therm-FM: Foundation Model is All You Need for 3D-ICs Thermal Simulation

Researchers introduced Therm-FM, a foundation-model-based neural operator for 3D-IC thermal simulation.

The method adapts a pretrained PDE model using multi-fidelity training, combining cheap approximate runs with limited high-fidelity calibration.

It outperformed prior approaches on public HotSpot and industrial 3D-IC package benchmarks for both steady-state and transient heat prediction.

By reducing data and simulation cost, Therm-FM could speed thermal-aware chip design and improve reliability in dense 3D integrated circuits.