README.md

---
license: apache-2.0
task_categories:
- tabular-regression
tags:
- cfd
- scientific-machine-learning
- pressure-prediction
- point-cloud
- submarine
- benchmark
pretty_name: Suboff Dataset
size_categories:
- 1K<n<10K
---

# Suboff Dataset

This public dataset contains the Suboff training split and the public validation inputs used by the leaderboard. It is intended for model training, local inference, and end-to-end leaderboard submission.

The data represents steady-state surface pressure over parameterized submarine hull geometries. Each geometry is evaluated under multiple Reynolds-number conditions. The model input is an irregular surface point cloud plus flow-condition features; the prediction target is pressure or pressure coefficient at each surface point.

## Dataset Contents

```text
data/
+-- train/
    +-- <geometry-id>-<condition-id>/
        +-- all_zones_combined.npy
+-- validation_input/
    +-- <geometry-id>-<condition-id>/
        +-- all_zones_combined.npy
```

This split contains:

| Split | Label availability | Geometries | Conditions per geometry | Files |
| --- | --- | ---: | ---: | ---: |
| Train | Public pressure labels | 158 | 6 | 948 |
| Validation input | No pressure labels | 39 | 6 | 234 |

Training `.npy` files are approximately 8 MB each. Validation-input `.npy` files are smaller because the pressure column has been removed.

## File Format

Training sample files are NumPy arrays:

```text
all_zones_combined.npy: float64 array with shape [N, 5]
```

The five columns are:

| Column | Name | Description |
| ---: | --- | --- |
| 0 | `x` | Surface point x-coordinate |
| 1 | `y` | Surface point y-coordinate |
| 2 | `z` | Surface point z-coordinate |
| 3 | `pressure` | CFD pressure value at the surface point |
| 4 | `zone_id` | Integer-like surface zone identifier |

Validation input files are NumPy arrays:

```text
all_zones_combined.npy: float64 array with shape [N, 4]
```

The four validation-input columns are:

| Column | Name | Description |
| ---: | --- | --- |
| 0 | `x` | Surface point x-coordinate |
| 1 | `y` | Surface point y-coordinate |
| 2 | `z` | Surface point z-coordinate |
| 3 | `zone_id` | Integer-like surface zone identifier |

The point count `N` can vary by sample. These files are irregular surface point clouds and should not be treated as structured grids.

## Flow Conditions

The condition ID is encoded in the folder name suffix, for example:

```text
hull-tail-fuyi-r-l005-r01-sail-bottom-tail-1p3-top-h001-3p97e7
```

Here `3p97e7` denotes the flow condition. The current loader maps these condition IDs to Reynolds numbers and inlet velocities:

| Condition ID | Reynolds number | Velocity |
| --- | ---: | ---: |
| `1p32e7` | 13,200,000 | 3.0507 |
| `2p23e7` | 22,300,000 | 5.1444 |
| `2p64e7` | 26,400,000 | 6.0962 |
| `3p10e7` | 31,000,000 | 7.1611 |
| `3p57e7` | 35,700,000 | 8.2311 |
| `3p97e7` | 39,700,000 | 9.1520 |

## Model Input and Target

A typical model consumes:

```text
point features: [x, y, z, zone_id]
global features: [reynolds_number, velocity]
```

The training target can be the raw pressure column or the pressure coefficient:

```text
Cp = pressure / (0.5 * rho * velocity^2)
```

The reference loader currently uses:

```text
rho = 998.2
```

## Loading Example

```python
from pathlib import Path
import re
import numpy as np

ROOT = Path("data/train")

REYNOLDS_TO_VELOCITY = {
    "1p32e7": 3.0507,
    "2p23e7": 5.1444,
    "2p64e7": 6.0962,
    "3p10e7": 7.1611,
    "3p57e7": 8.2311,
    "3p97e7": 9.1520,
}


def parse_condition(sample_dir_name: str):
    match = re.search(r"-(\d+p\d+e\d+)$", sample_dir_name)
    if match is None:
        raise ValueError(f"Cannot parse condition from {sample_dir_name}")
    condition_id = match.group(1)
    reynolds = float(condition_id.replace("p", "."))
    velocity = REYNOLDS_TO_VELOCITY[condition_id]
    return condition_id, reynolds, velocity


sample_file = next(ROOT.glob("*/all_zones_combined.npy"))
sample_name = sample_file.parent.name
condition_id, reynolds, velocity = parse_condition(sample_name)

array = np.load(sample_file)
xyz = array[:, :3].astype("float32")
pressure = array[:, 3].astype("float32")
zone_id = array[:, 4].astype("int64")

rho = 998.2
cp = pressure / (0.5 * rho * velocity**2)

print(sample_name)
print(array.shape, array.dtype)
print(condition_id, reynolds, velocity)
print(xyz.shape, pressure.shape, zone_id.shape, cp.shape)
```

## Leaderboard Submission Format

Download the files under `data/validation_input/`, run local inference, and submit a ZIP file to the Space:

```text
submission.zip
+-- predictions/
    +-- <geometry-id>-<condition-id>.npy
```

Each prediction file must contain predicted pressure coefficient values:

```text
pred_cp: float array with shape [N] or [N, 1]
```

The point order must match the corresponding `data/validation_input/<sample-id>/all_zones_combined.npy` file exactly.

## Public and Private Data Boundary

This repository is public. It contains training files with pressure values and validation-input files without pressure values. The private validation-truth files are stored separately in `s3no-benchmark/suboff_val` and should not be copied into this public repository.

For a leaderboard workflow:

- Public train data can include coordinates, zone IDs, condition features, and pressure/Cp labels.
- Private validation or test truth should include labels but remain inaccessible to participants.
- Public validation/test inputs should omit pressure/Cp labels if they are used for hidden scoring.

## License

This dataset repository is released under the Apache License 2.0 as indicated in the Dataset Card metadata.

If you use this data in a benchmark, report the exact split and the condition mapping above so results remain reproducible.