File size: 2,599 Bytes
10ab127
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
---
license: apache-2.0
library_name: pytorch
tags:
- OneScience
- neural-operator
- darcy-flow
---

# SirenFNO: Darcy 32x32

This package reproduces the full-spectrum SirenFNO experiment for two-dimensional Darcy flow in Table 2 of the [SirenFNO paper](https://ijcai.org/proceedings/2026/0535.pdf). 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

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

The original 32x32 Darcy tensors are published by the [NeuralOperator team on Zenodo](https://zenodo.org/records/12784353). The same two files are mirrored for this experiment in the [OneScience/SirenFNO-Darcy32 ModelScope dataset](https://www.modelscope.cn/datasets/OneScience/SirenFNO-Darcy32). No dataset files are included in this Hugging Face model repository.

```bash
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.

```bash
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.