PseudoTimePINN-BackwardFacingStep

Model Introduction

This package reproduces the Re=800 backward-facing-step experiment from When PINNs Go Wrong: Pseudo-Time Stepping Against Spurious Solutions. An adaptive pseudo-time PINN predicts steady velocity and pressure from (x, y). The SU2 reference field is used for evaluation only, not for supervised parameter updates.

Model Description

The network uses a Fourier-feature MLP and a SOAP optimizer. The included run completed 100,000 training steps on one DCU. weight/best_model.pth was selected at step 47,200 using an unlabeled physics validation score; weight/last_model.pth is retained because the default inference compares the best and final checkpoints.

Intended Use

Task	Description
Backward-facing-step flow	Approximate the separated Re=800 flow field
PINN evaluation	Compare predictions with the author's SU2 reference field

Usage

Download the Model Package

hf download OneScience-Group/PseudoTimePINN-BackwardFacingStep --local-dir ./PseudoTimePINN-BackwardFacingStep
cd PseudoTimePINN-BackwardFacingStep

Install Dependencies

Use an environment with a compatible PyTorch GPU/DCU build. On the tested DCU platform, activate DTK and a compatible Conda environment, then install requirements.txt and modelscope[datasets]. Full training requires an allocated Slurm compute node; the provided configuration does not silently fall back to CPU.

pip install -r requirements.txt
pip install 'modelscope[datasets]'

Evaluation Data

The only external file is the author's 31,265-point SU2 reference field. Download it from the OneScience/BackwardFacingStep-SU2-Re800 dataset into the package's data/ directory:

modelscope download --dataset OneScience/BackwardFacingStep-SU2-Re800 --local_dir ./data

The expected file is data/flow.vtu, SHA256 004f10b2b9ce1e38542537da648c1221730308310f1f7fa675dc7a92b4900d8a. PDE and boundary training points are generated by the code.

Train

From the package root on an allocated GPU/DCU compute node:

python scripts/train.py

The configured full run has 100,000 steps. Save copies of the bundled checkpoints before retraining.

Inference and Results

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

The best checkpoint's measured relative L2 error in u was 0.003582, compared with 0.002320 in the paper's Table 1. The final checkpoint was substantially worse (0.093482); use the best checkpoint for applications. The comparison is one run, not a claim of exact paper-level reproduction.

OneScience Repositories

Platform	Main repository	Skills repository
Gitee	OneScience	OneSkills
GitHub	OneScience	OneSkills

Citation and License

• Paper: When PINNs Go Wrong.
• Reference field: JAXPI2 backward-facing-step data. The data is distributed separately on ModelScope; consult the original repository's applicable terms.