# Inference example

This folder demonstrates how to run FuXi-CFD ONNX inference end-to-end:
- load raw inputs (`data/inputs.npz`)
- normalize + resample to model input grid
- run ONNX with onnxruntime
- de-normalize model outputs back to physical scale

## Run

```bash
python scripts/infer.py \
  --model ../model/fuxicfd_model.onnx \
  --input data/inputs.npz \
  --output data/prediction.npz
```

## Input format

`data/sample_input.npy` stores a Python dict (load with `allow_pickle=True`) containing:
- `u_100m, v_100m`: shape `(9, 9)`
- `dem, roughness`: shape `(301, 301)`

The final model input tensor is built as `(4, 300, 300)` with channel order:
`[u_100m, v_100m, dem, roughness]`

## Output format

`data/prediction.npz` contains `u, v, w, k`, each of shape `(27, 300, 300)`.

## Normalization
See `normalization/README.md` for exact variable orders and shapes.