day3/IC-ThermBench-S2-S5

IC-ThermBench: S2–S5 datasets

Steady-state thermal fields for Scopes S2–S5 of IC-ThermBench, an open, progressive benchmark for generalizable 2.5D/3D-IC thermal learning.

• Paper: https://arxiv.org/abs/2608.23977
• Code, splits, and evaluation: https://github.com/Day333/ThermalBench

Contents

level2_steady/   input.mat  output.mat                  P=3   15,000  Cases 1–10 × 1,500
level3_steady/   input.mat  output.mat                  P=4   15,000  Cases 1–10 × 1,500
level4_steady/   input.mat  output.mat                  P=7   15,000  Cases 1–10 × 1,500
level5_steady/   input.mat  output.mat  manifest.json   P=7    5,000  Cases 11–15 × 1,000

Folder	Paper scope	Input channels (in order)
level2_steady	S2: layout variation	chiplet_power, grid_x, grid_y
level3_steady	S3: + material variation	S2 + local_thermal_k
level4_steady	S4: + boundary-condition variation	S3 + ambient_K, h_w_m2k, r_convec_k_per_w
level5_steady	S5: unseen package systems (evaluation only)	same as S4

output.mat holds the reference temperature field in kelvin on a 64 × 64 grid. level5_steady/manifest.json maps each S5 sample to its case. The train/validation/test splits are fixed by index in the code repository; see docs/DATASETS.md for the tensor layout and split details.

Usage

Download into the datasets/ folder at the root of the code repository:

hf download day3/IC-ThermBench-S2-S5 --repo-type dataset --local-dir datasets
python run.py --model UFNO --data level2 --task test

License

The S2–S5 datasets and the S5 manifest are released under CC BY 4.0. See LICENSE-DATA for attribution terms.

Citation

@article{icthermbench2026,
  title   = {IC-ThermBench: An Open, Progressive Benchmark for Generalizable
             2.5D/3D-IC Thermal Learning},
  author  = {David Hang and Wenkai Yang and Kuiye Ding and Haiyang Xin and Jacky Wei},
  journal = {arXiv preprint arXiv:2608.23977},
  year    = {2026},
  url     = {https://arxiv.org/abs/2608.23977}
}