---
license: mit
pretty_name: ResBench reference
tags:
- geology
- reservoir-modeling
- benchmark
- 3d
---

# ResBench reference

The reference data that [ResBench](https://github.com/SciLM-ai/ResBench) scores
generative reservoir models against. Every volume was produced by the ResMill
process simulator (commit in `ENGINE.txt`) under the same rules as the
[SiliciclasticReservoirs](https://huggingface.co/datasets/SciLM/SiliciclasticReservoirs)
training dataset. You do not need to download it by hand:

```bash
pip install "resbench[download]"
resbench download        # fetches this repository into the local cache
resbench score SUBMISSION --out results.json
```

## Contents

| path | what |
|---|---|
| `manifest.csv` | one row per reference item: task, environment, id, conditioning parameters, well position |
| `volumes/<env>/volumes.npz` | 512 test-split 64 x 64 x 32 facies windows per environment, regenerated bit for bit from the dataset rows |
| `repeats/<env>/cond0-4.npz` | 256 ResMill runs of each of five fixed parameter rows (unconditional repeats) |
| `repeats/<env>/well1-5.npz` | five well conditions per environment, each an ensemble of windows whose centre column matches the well exactly |
| `fields/<env>/fields.npz` | 32 field-scale volumes per environment (lobe and delta 512 x 512 x 32, channels 512 x 128 x 32), each the dataset-rule 32-cell window of a 64-cell engine column |
| `ENGINE.txt` | the ResMill commit every volume came from |

Eight environments: lobe, delta and six channel presets (PV_SHOESTRING,
CB_LABYRINTH, CB_JIGSAW, SH_DISTAL, SH_PROXIMAL, MEANDER_OXBOW). Volumes are
binary sand (1) / mud (0), `int8`.

Everything here can be rebuilt from the ResBench repository's `tools/`
(command sequence in its SPEC.md) and ResMill at the commit in `ENGINE.txt`.
Scored against itself, this reference matches in all 31 (task, check) cells.