ScireX Transolver · Chip cooling

A trained Transolver neural operator from the open-source scirex library (JAX/Flax) for the Chip cooling benchmark. The weights are the last-epoch weights of the run described below and are stored as model.safetensors.

Results

Lower is better, except for Spearman correlations. Figures are those of the ScireX results table and benchmark pages: benchmark page and the GitHub README. This is the full-normalisation variant of the Transolver run. The training-time column is what was recorded for it when it was trained. The re-evaluated column was obtained by loading this repo's config.json (its normalisation statistics pass infer.py's check against the benchmark data) and running scripts/benchmarks/chip_cooling/infer.py on the 400 test samples (CPU). These are the values you should expect to reproduce.

Inputs and outputs

Input: 6 boundary-condition parameters of a dual-chip cold plate. Output: steady-state temperature (K) on a 2D slice (11,110 mesh nodes). 1600 training and 400 test samples. The normalisation statistics are in config.json; the weights are unusable without them.

Training data

Trained on the data of Curl & Hu (2026), Dryad doi:10.5061/dryad.k0p2ngfp5, released under CC0 1.0.

Training

Transolver, 1,807,561 parameters. Adam with weight decay 1e-6, gradient clipping 1.0, LR 1e-4 with 50 warm-up epochs then cosine to 1e-6, batch 4. All 11,110 nodes; 5 layers, width 256, 8 heads, 32 slices. Temperature z-scored with one scalar, parameters z-scored per parameter (training statistics only). Relative L2 loss, 500 epochs, seed 42, final-epoch weights. About 1.8M trainable parameters per model. Training time: see the GitHub docs page.

Full setup: doc/benchmarks/chip_cooling/results/comparative.md.

How to use

hf download zenteiq-ai/scirex-transolver-chip-cooling --local-dir ckpt/scirex-transolver-chip-cooling

Then run the benchmark's scripts/benchmarks/chip_cooling/infer.py from the GitHub repository. The exact command is under Reproduce below. Check your install against sample.safetensors, which holds one test input (input) and the model's expected output (expected_output). config.json holds the model settings and the normalisation statistics.

Reproduce

From a checkout of the GitHub repository at tag v1.0.0:

hf download zenteiq-ai/scirex-transolver-chip-cooling --local-dir ckpt/scirex-transolver-chip-cooling
python scripts/benchmarks/chip_cooling/infer.py \
    --checkpoint ckpt/scirex-transolver-chip-cooling \
    --run transolver \
    --data-dir $DATA_ROOT/chip_cooling

$DATA_ROOT is the folder that holds the benchmark datasets (see reproducibility.md).

Results table for this benchmark: comparative.md.

Limitations

Valid only within the trained input ranges of the benchmark data. Not for safety-critical design without separate validation. Bug reports: GitHub Issues.

Other ScireX checkpoints

Same benchmark, other models:

Same model, other benchmarks:

Licence and citation

The weights are released under the Apache License 2.0 (see LICENSE; dataset credit in NOTICE). The code is LGPL-3.0. Cite the software through the CITATION.cff in the GitHub repository and the architecture paper:

• Transolver: Wu et al., ICML 2024, arXiv:2402.02366