README.md

---
license: apache-2.0
pretty_name: BEvAn — COSI activation simulations (β⁺ / 511 keV)
tags:
  - astrophysics
  - gamma-ray
  - compton-telescope
  - megalib
  - cosi
  - simulation
size_categories:
  - 1K<n<10K
---

# BEvAn — COSI activation simulations

MEGAlib simulation files backing [`cosi-betadecay/BEvAn`](https://github.com/cosi-betadecay/BEvAn),
a physics-based likelihood classifier that tags each simulated COSI event as a
β⁺ / positron-annihilation signal (back-to-back 511 keV photons) or background
from its Compton kinematics.

All events come from a **cosima activation simulation** (`BeamType Activation`,
`DecayMode ActivationBuildup`): cosmic-ray protons irradiate the instrument, the
activated isotopes decay, and the β⁺ emitters among them produce the 511 keV
annihilation signal the classifier is trained to find. The three-step cosima
source files that generate this are in the GitHub repo under
[`data/cosima/`](https://github.com/cosi-betadecay/BEvAn/tree/main/data/cosima)
and are **not** duplicated here. Produced with MEGAlib 4.00.00.

## Datasets

One folder per detector geometry. The geometries ship with MEGAlib under
`$MEGALIB/resource/examples/geomega/`:

| Folder | Geometry (`.geo.setup`) | Size |
|---|---|---|
| `SPILike/` | `special/SPILike` | 382 MB |
| `NCT/` | `special/Simple_Ge_NCT_try1_040721` | 848 MB |
| `Max/` | `special/Max` | 1.2 GB |
| `GeACT/` | `special/GeACT` | 3.8 GB |
| `COSIBalloon_9det/` | `cosiballoon/COSIBalloon.9Detector` | 1.0 GB |
| `COSIBalloon_10det/` | `cosiballoon/COSIBalloon.10Detector` | 980 MB |
| `COSIBalloon_12det/` | `cosiballoon/COSIBalloon.12Detector` | 1.1 GB |

## Files in each folder

For a dataset `{name}`:

| File | What it is |
|---|---|
| `{name}.sim` | **Concatenation file written by `mcosima` — it holds no events.** It is a short header plus `IN {name}.incN.id1.sim.gz` lines pointing at the chunks below. This is the path you pass to the pipeline; MEGAlib follows the pointers. |
| `{name}.incN.id1.sim.gz` | The actual simulated events, one chunk per parallel `mcosima` thread (N = 1…4). Required — `{name}.sim` is useless without them. |
| `{name}.tra` | Compton-reconstructed events (`revan` output) for `{name}.sim`. |
| `{name}_P1.*`, `{name}_P2.*` | Two **dedicated prior simulations** against the same geometry, each a full `.sim` (+ `.tra`) set of its own. Their ANNI-labeled events, counted per hit-multiplicity bucket, form the classifier's class priors. The training split is never used for the prior, which is why these are separate runs. Passed via `--prior-sim`. `_P1` files are self-contained `.sim` files from an earlier run; `_P2` files use the same `mcosima` concatenation layout as above. |

The `COSIBalloon_*` folders additionally carry a `crossections/` directory —
MEGAlib-generated cross-section tables. They are regenerated on demand by
MEGAlib and are included only so the folders are a byte-exact mirror.

## Results

`results/` holds the **outputs** of the ablation study over the datasets above —
a mirror of `ablations/results/` in the GitHub repo (which remains the source of
truth; these are generated files, published here so the figures and numbers are
citable alongside the inputs they came from). 25 MB total.

One folder per run, named by timestamp:

| Run | Contents |
|---|---|
| `2026-07-14_09-33-32/` | includes a `no_calibration` ablation, since removed |
| `2026-07-15_00-45-23/` | |
| `2026-07-15_05-06-29/` | most recent; adds `deployment` and `label_window` |

Each run folder has:

- `tables/` — one CSV per ablation (`factor_contributions`, `learned_weights`,
  `no_ckd_order`, `gt_tolerance` / `label_window`, …) plus a `summary.csv`.
  `deployment.csv` is the champion deployed per dataset at the dedicated-prior
  operating point.
- `figures/` — one PNG per ablation at the top level, plus a per-dataset
  subfolder (`SPILike/`, `NCT/`, `Max/`, `GeACT/`, `COSIBalloon_*/`) of the
  per-run plots for that geometry.

Regenerate with `python ablations/main.py` from the GitHub repo.

## Download

```bash
hf download aryaraeesi/BEvAn-data --repo-type dataset --local-dir data/
```

Note this pulls `results/` into `data/` too. For just the simulation inputs, or
a single dataset (remember to take its `.inc*` chunks along with the `.sim`):

```bash
hf download aryaraeesi/BEvAn-data --repo-type dataset \
  --include "SPILike/*" --local-dir data/
```

Or just the results, without the multi-GB simulations:

```bash
hf download aryaraeesi/BEvAn-data --repo-type dataset \
  --include "results/*" --local-dir .
```

## Gotcha: the geometry path is absolute

The `Geometry` line inside every `.sim` header is an **absolute path from the
machine that ran the simulation** (`/home/arya/Documents/megalib/...`). MEGAlib
will not find the geometry on your machine unless you either point that line at
your own `$MEGALIB` checkout or pass the geometry explicitly — which is what the
BEvAn entry scripts do via `--geo-file`:

```bash
python src/BEvAn/analysis.py \
  --geo-file $MEGALIB/resource/examples/geomega/special/SPILike.geo.setup \
  --sim-file data/SPILike/SPILike.sim \
  --tra-file data/SPILike/SPILike.tra \
  --prior-sim data/SPILike/SPILike_P1.sim data/SPILike/SPILike_P2.sim
```

## License

Apache-2.0, matching the BEvAn repository.