---
license: cc-by-4.0
tags:
  - physics
  - pde
  - neural-operator
  - fluid-dynamics
  - combustion
  - benchmark
datasets:
  - AI4Science-WestlakeU/RealPDEBench
---

# RealPDEBench Model Checkpoints

Trained model checkpoints for [RealPDEBench](https://github.com/AI4Science-WestlakeU/RealPDEBench), a benchmark for evaluating neural PDE solvers on real-world experimental data.

## Models (10 architectures)

| Model | Type | File Size (per checkpoint) |
|-------|------|---------------------------|
| DPOT-L | Transformer | 2.5-2.6 GB |
| FNO | Spectral | 385M-2.1G |
| Galerkin Transformer | Transformer | 386-642M |
| WDNO | Diffusion | 351M-1.4G |
| DPOT-S | Transformer | 118-159M |
| U-Net | CNN | 88-89M |
| CNO | Hybrid | 31M |
| MWT | Wavelet | 22M |
| Transolver | Transformer | 17M |
| DeepONet | Neural Operator | 14M |

## Scenarios (5)

| Scenario | Description |
|----------|-------------|
| `cylinder` | Flow past a circular cylinder |
| `controlled_cylinder` | Actively controlled cylinder flow |
| `fsi` | Fluid-structure interaction |
| `foil` | Flow past an airfoil |
| `combustion` | Turbulent combustion |

## Training Paradigms

| File | Paradigm |
|------|----------|
| `numerical.pth` | Trained on numerical simulation data only |
| `real.pth` | Trained on real experimental data only |
| `finetune.pth` | Pretrained on numerical, finetuned on real |
| `numerical_base_for_finetune.pth` | Numerical pretrain base (DPOT-S/L only) |

<details>
<summary><b>Why DPOT-S/L include <code>numerical_base_for_finetune.pth</code></b></summary>

For DPOT-S/L, `numerical.pth` and `numerical_base_for_finetune.pth` are two snapshots from the same numerical pretraining run. `numerical.pth` is the best-val checkpoint, kept for numerical-only evaluation. `numerical_base_for_finetune.pth` is the last checkpoint of that same run, used as the starting point for finetuning.

DPOT is a foundation model carrying dataset-level bias from its original pretraining data. The best-val checkpoint of the numerical stage is premature for use as a finetune starting point, and finetuning from a premature checkpoint defeats the purpose of the numerical pretraining stage.

For all other (from-scratch) models, `numerical.pth` (best-val) serves both as the evaluation checkpoint and as the finetune starting point, so no extra file is needed.

</details>

## Directory Structure

```
{scenario}/{model}/{paradigm}.pth
configs/{scenario}/{model}.yaml
```

Example: `cylinder/fno/finetune.pth` + `configs/cylinder/fno.yaml`

## Quick Start

### Install

```bash
git clone https://github.com/AI4Science-WestlakeU/RealPDEBench.git
cd RealPDEBench && pip install -e .
```

### Download a Single Checkpoint

```python
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="AI4Science-WestlakeU/RealPDEBench-models",
    filename="cylinder/fno/finetune.pth",
)
```

### Download All Checkpoints for a Scenario

```python
from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="AI4Science-WestlakeU/RealPDEBench-models",
    allow_patterns="cylinder/**",
    local_dir="./checkpoints",
)
```

### Evaluate

```bash
python eval.py --config configs/cylinder/fno.yaml \
    --checkpoint_path ./checkpoints/cylinder/fno/finetune.pth \
    --dataset_type real --test_mode all
```

## Checkpoint Format

```python
checkpoint = torch.load("cylinder/fno/finetune.pth")
# Keys: model_state_dict, train_losses, val_losses,
#        iteration, best_iteration, best_val_loss
```

## DPOT Pretrained Weights

DPOT models require pretrained backbone weights (**not included here**).
Download via:

```bash
# Option 1: Built-in download script
python -m realpdebench.utils.dpot_ckpts_dl

# Option 2: From HuggingFace directly
# https://huggingface.co/hzk17/DPOT
```

## Dataset

The corresponding dataset is hosted at: [AI4Science-WestlakeU/RealPDEBench](https://huggingface.co/datasets/AI4Science-WestlakeU/RealPDEBench)


## Citation

If you find our work and/or our code useful, please cite us via:

```bibtex
@inproceedings{hu2026realpdebench,
      title={RealPDEBench: A Benchmark for Complex Physical Systems with Real-World Data}, 
      author={Peiyan Hu and Haodong Feng and Hongyuan Liu and Tongtong Yan and Wenhao Deng and Tianrun Gao and Rong Zheng and Haoren Zheng and Chenglei Yu and Chuanrui Wang and Kaiwen Li and Zhi-Ming Ma and Dezhi Zhou and Xingcai Lu and Dixia Fan and Tailin Wu},
      booktitle={The Fourteenth International Conference on Learning Representations},
      year={2026},
      url={https://openreview.net/forum?id=y3oHMcoItR},
      note={Oral Presentation}
}
```

## License

CC BY 4.0