MoNo-light: Darcy 85x85

This package reproduces the MoNo-light Darcy experiment in Table 3 of the MoNo paper. It predicts pressure from permeability on an 85x85 grid using multiscale latent variables and Sinkhorn transport. The single-seed run completed 500 epochs with 1,000 training and 200 test samples.

Results

The reloaded final-epoch checkpoint achieves mean test relative L2 0.00693725, versus 0.0070 in the paper, a numerical difference of about 0.9%. Hardware, software, and the Sinkhorn kernel differ from the authors' setup, so this is not evidence of strict same-condition superiority. Full metrics and plots are in results/.

Download

hf download OneScience-Group/MoNo-light --local-dir ./MoNo-light
cd MoNo-light

The two original 421x421 Darcy MAT files are also hosted in the OneScience/HNO-Darcy ModelScope dataset. This Hugging Face repository contains the model only, not the raw data. Download the dataset from the model directory:

modelscope download --dataset OneScience/HNO-Darcy --local_dir ./data

The package reads data/piececonst_r421_N1024_smooth1.mat and data/piececonst_r421_N1024_smooth2.mat. Both contain coeff and sol arrays; the loader uses the first 1,000 and 200 samples respectively and strides the 421x421 grid by five. It creates a local cache on first use.

Environment and usage

Use Python 3.11, PyTorch on an allocated GPU/DCU, onescience[cfd-gpu] or onescience[cfd-dcu], SciPy, PyYAML, and Matplotlib. The raw MAT files total about 3.39 GB; allow sufficient RAM for the first load.

python scripts/inference.py
python scripts/result.py

For a new 500-epoch training run, move the included weight/ and results/ outside the model directory before running python scripts/train.py. The published inference accepts a relocated data directory while retaining the model/training-configuration check. The weights are unchanged from the original run.

Checkpoint: weight/best_model.pth (SHA256 d20393f9ae89d38f53b8da9e09799db57bef37154ec8c7fddb9a9dd4a6666122). Original data source: FNO Darcy files. Check the source's terms before redistributing the data.

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

Paper for OneScience-Group/MoNo-light