ScireX
Collection
Trained scientific machine learning models from ScireX, ZenteiQ’s open-source library. • 33 items • Updated • 1
You need to agree to share your Hugging Face information to access this model
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
These weights are licensed under CC BY-NC 4.0 for non-commercial use. For commercial use, contact research@zenteiq.ai.
Log in or Sign Up to review the conditions and access this model content.
ScireX Transolver · Pipe flow
A trained Transolver neural operator from the open-source scirex library (JAX/Flax) for the Pipe flow benchmark. The weights are the last-epoch weights of the run described below and are stored as model.safetensors.
Results
| Metric | This model | Published reference |
|---|---|---|
| Test relative L2, this checkpoint (tuned recipe, lowest test error of 3 seeds) | 0.386% | 0.33% ± 0.02% (Wu et al., ICML 2024) |
| Test relative L2, default recipe (seed 0, for comparison) | 0.523% | – |
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 repository holds the tuned-recipe checkpoint, the lowest test error of three seeds (0.446%, 0.386%, 0.387%; mean 0.406% ± 0.034%). The tuned recipe and the choice of the best seed both use the test set, as there is no validation split. For context, Geo-FNO reports 0.67% on this benchmark.
Inputs and outputs
Input: mesh coordinates of a deformed pipe (the centreline varies per sample). Output: horizontal velocity on the mesh. The normalisation statistics are in config.json; the weights are unusable without them.
Training data
Trained on the Geo-FNO pipe data of Li et al. (JMLR 2023).
Training
Transolver, 3,073,985 parameters. AdamW, weight decay 1e-5, gradient-norm clip 0.1, tuned recipe. Input coordinates and targets unit-Gaussian normalised with training statistics. Cosine one-cycle LR (peak 1e-3), batch 8, 500 epochs, three seeds; last-epoch weights, no epoch chosen by test error. There is no validation split. Training time: 106 to 215 min per seed (partly on a shared GPU).
Full setup: doc/benchmarks/pipe/results/comparative.md.
How to use
hf download zenteiq-ai/scirex-transolver-pipe --local-dir ckpt/scirex-transolver-pipe
Then run the benchmark's scripts/benchmarks/pipe/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-pipe --local-dir ckpt/scirex-transolver-pipe
python scripts/benchmarks/pipe/infer.py \
--checkpoint ckpt/scirex-transolver-pipe \
--run transolver_tuned \
--data-dir $DATA_ROOT/pipe
$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:
| Model | Repo |
|---|---|
| FNO | zenteiq-ai/scirex-fno-pipe |
| GINO | zenteiq-ai/scirex-gino-pipe |
| WNO | zenteiq-ai/scirex-wno-pipe |
Same model, other benchmarks:
| Benchmark | Repo |
|---|---|
| Darcy flow 2D (data-driven) | zenteiq-ai/scirex-transolver-darcy |
| Darcy flow 2D (physics-only, extra-fields run) | zenteiq-ai/scirex-transolver-darcy-physics-extra-fields |
| Navier-Stokes 2D | zenteiq-ai/scirex-transolver-navier-stokes |
| Chip cooling | zenteiq-ai/scirex-transolver-chip-cooling |
| AirfRANS | zenteiq-ai/scirex-transolver-airfrans |
| Car-Design-ShapeNetCar | zenteiq-ai/scirex-transolver-shapenetcar |
| Car-CFD | zenteiq-ai/scirex-transolver-car-cfd |
Licence and citation
The weights are released under CC BY-NC 4.0 (see LICENSE; dataset credit in NOTICE). The code is LGPL-3.0. For commercial use, contact research@zenteiq.ai. Cite the software through the CITATION.cff in the GitHub repository and the architecture paper:
- Transolver: Wu et al., ICML 2024, arXiv:2402.02366
- Downloads last month
- 1
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support