CFD Model Develop

cfd fno models run, also quckliy download code

Install

openclaw skills install @onescience-ai/cfd-model-run

FNO Custom Dataset Training Assistant

Description

Use this skill whenever the user wants to train or adapt a Fourier Neural Operator (FNO) model on their own dataset, perform neural PDE operator training, or build a surrogate model for structured grid data.

This skill downloads the FNO codebase from ModelScope, adapts it to the user's dataset, and launches training and inference.

The first execution automatically downloads the FNO code from ModelScope and caches it locally.

Use this skill when the user asks to

Train FNO on custom data
Adapt Fourier Neural Operator to a new dataset
Build an AI surrogate model for PDE solving
Fine-tune or retrain FNO
Use FNO for structured grid prediction
Predict time-series flow fields with neural operators
Replace CFD solvers with a learned operator

Workflow

Step 1. Download Model Code

Check whether the FNO codebase already exists locally. If not, use modelscope.snapshot_download with model_id="OneScience/FNO_B" to download it and print the local path.

Step 2. Verify Runtime Environment

Required dependencies: Python >= 3.10, PyTorch, scipy, ruamel.yaml, tqdm, matplotlib, numpy.

Install any missing packages with pip. CUDA is optional — the code auto-detects cuda:0 and falls back to CPU.

Step 3. Understand Expected Data Format

Ask the user about their data — shape, dimensionality (1D / 2D / 3D grid), number of physical channels, and time steps. The model expects batches with three keys:

Key	Shape	Description
`pos`	`(B, N, space_dim)`	Grid point coordinates
`x`	`(B, N, t_in × out_dim)`	Input time window (flattened grid)
`y`	`(B, N, t_out × out_dim)`	Prediction target (flattened grid)

where N is the total number of grid points (H for 1D, H*W for 2D, H*W*D for 3D).

Step 4. Adapt the Data Pipeline

Modify NavierStokesDataset._init_data() in navier_stokes.py to load the user's data and populate self.x, self.y, self.pos, and self.spatial_shape with the appropriate tensors and shape tuple.

Then update conf/fno.yaml under datapipe.source to point to the user's data directory and file, and update datapipe.data fields — ntrain, ntest, t_in, t_out, out_dim — to match the actual dataset.

Step 5. Extend to 1D or 3D Grids (if needed)

fno_layers.py only ships SpectralConv2d. For other dimensionalities, the user needs to implement the corresponding spectral convolution class and update model.py.

1D grid: Implement SpectralConv1d in fno_layers.py using torch.fft.rfft / irfft with a single weight tensor of shape (in_channels, out_channels, modes1). In model.py, replace SpectralConv2d with SpectralConv1d, replace nn.Conv2d pointwise layers with nn.Conv1d, and reshape the feature tensor to (B, C, L) in the forward pass instead of (B, C, H, W).

3D grid: Implement SpectralConv3d in fno_layers.py using torch.fft.rfftn / irfftn along the last three dimensions with four weight tensors covering the eight frequency corners. In model.py, replace spectral and pointwise layers with their 3D equivalents and reshape to (B, C, D, H, W) in the forward pass.

In both cases, also update conf/fno.yaml: set model.space_dim to 1 or 3, adjust modes1/modes2/modes3 to values that do not exceed half the grid size in each dimension.

Step 6. Launch Training

Run train.py from the downloaded model directory. The loop evaluates val_rel_l2 and val_mse every eval_interval epochs, saves the best checkpoint as best_model.pt under training.output_dir, and early-stops after patience epochs without improvement.

Key parameters in conf/fno.yaml:

Parameter	Default	Description
`training.num_epochs`	500	Max training epochs
`training.lr`	0.001	AdamW learning rate
`training.teacher_forcing`	true	Use ground truth as next-step input
`model.hidden_channels`	64	Hidden feature channels
`model.num_layers`	4	Fourier layer depth

Step 7. Run Inference and Visualization

Run inference.py from the model directory. Set the FNO_CHECKPOINT environment variable to override the checkpoint path from the config. Outputs are saved to vis_results/ under training.output_dir: a .pt file with pos, input, prediction, target tensors, and a .png comparison plot of Target / Prediction / Abs Error.

Trigger Keywords

FNO

Fourier Neural Operator

Neural PDE

PDE surrogate

Structured grid prediction

Time series flow

Train FNO

Custom dataset FNO

Neural operator training

1D FNO

3D FNO

AI surrogate model

Scientific AI

AI4S

Notes

The codebase is fully standalone — no onescience repository installation required.
Single-device only (single GPU or CPU); DDP is not supported.
On first run, ModelScope downloads the full codebase. Subsequent runs reuse the cache.
If download fails, set MODELSCOPE_API_TOKEN or use a mirror endpoint.
For 1D or 3D extension, SpectralConv1d / SpectralConv3d do not exist in the downloaded code and must be implemented by the user in fno_layers.py.