CROP-NavierStokes-HighRe

Model Introduction

This package reproduces the CROP single-step vorticity prediction experiment for two-dimensional Navier-Stokes flow at Re=10000. It maps the field at time index 0 to time index 5 on a 64 x 64 grid.

Paper: Discretization-invariance? On the Discretization Mismatch Errors in Neural Operators

Model Description

The model uses band-limited lifting and projection around a multiscale local convolutional operator. Input and output tensors have shape [batch,1,64,64].

Usage

Download the Model Package

hf download OneScience-Group/CROP-NavierStokes-HighRe --local-dir ./CROP-NavierStokes-HighRe
cd CROP-NavierStokes-HighRe

Install the Runtime

Use a Python 3.11 environment with the appropriate GPU/DCU PyTorch build, numpy, h5py, PyYAML, and matplotlib. For a OneScience DCU environment, install onescience[cfd-dcu]; for GPU, use onescience[cfd-gpu].

Training Data

The original HDF5 data comes from the CROP research repository. A provenance-preserving copy is available on ModelScope. Place new_ns_train_1e-4_T30_N2048_s64.h5 in ./data before running the scripts. The dataset is not mirrored to Hugging Face by this model repository.

Training

python scripts/train.py

The best checkpoint is written to weight/best_model.pth.

Inference

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

The included checkpoint has SHA256 962c9ed870d5c916614abf59471bb89633bde8b63891e9ff152848aca1624655. Its test mean relative L2 is 4.3402%; the paper reports 4.06% for this experiment.

Citation and License

Please cite the CROP paper. This reproduction package is released under Apache-2.0; the original dataset has no separately identified redistribution license.