SirenFNO: Darcy 32x32

This package reproduces the full-spectrum SirenFNO experiment for two-dimensional Darcy flow in Table 2 of the SirenFNO paper. The model maps a permeability field to its pressure solution. It has four spectral blocks, width 32, and was trained for 500 epochs on 1,000 samples.

Results

The reloaded checkpoint achieves a mean per-sample relative L2 error of 0.04524468 on 200 test fields. The paper reports 0.0351; this run is about 28.9% higher. The loss definition and some unspecified implementation details may differ, so this is not a strict same-condition comparison. The original training report and plots are in results/.

Download

hf download OneScience-Group/SirenFNO --local-dir ./SirenFNO
cd SirenFNO

The original 32x32 Darcy tensors are published by the NeuralOperator team on Zenodo. The same two files are mirrored for this experiment in the OneScience/SirenFNO-Darcy32 ModelScope dataset. No dataset files are included in this Hugging Face model repository.

modelscope download --dataset OneScience/SirenFNO-Darcy32 --local_dir ./data

Run the commands below from the model directory. The published configuration reads data/darcy_train_32.pt and data/darcy_test_32.pt and writes to results/. Each file contains float32 x (permeability) and y (pressure) tensors on a 32x32 grid. The inference script checks the data SHA256 against the checkpoint.

Environment and usage

Use Python 3.11, PyTorch with a supported GPU or DCU, onescience[cfd-gpu] or onescience[cfd-dcu], PyYAML, and Matplotlib. On a cluster, run model computation on an allocated accelerator node.

python scripts/inference.py --config config/config.yaml
python scripts/result.py --config config/config.yaml

To start a new 500-epoch training run, move the included weight/ and results/ outside the model directory first, then run python scripts/train.py --config config/config.yaml. The published inference accepts a different data/output location but retains the model-configuration and data-hash checks. The weights are unchanged from the original run.

Checkpoint: weight/best_model.pth (SHA256 ceb2392d5ba68e386f08e0f117ea1705df1f3a0185d3cef0971f07951113ce1e). The dataset's original use and redistribution terms should be checked at the Zenodo source.

Downloads last month
32
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support