---
license: apache-2.0
tags:
  - fusion
  - tokamak
  - diii-d
  - pedestal
  - onnx
library_name: onnx
---

# edensfit89

> **Consumer note.** Most users should not load this bundle's
> ONNX files directly. Use the `PedestalEnsemble` Python wrapper
> shipped in the [PedestalPredictor GitHub repo](https://github.com/SCS-Lab/PedestalPredictor),
> which loads all five bundles with one call and exposes a
> unified `predict_one(...)` API. Direct ONNX access documented
> below is for advanced users who want to integrate a single
> bundle into an existing ONNX-only pipeline.

## Summary

- **Task:** regression
- **Target:** `edens_ped`
- **Dataset version:** `v1_446` (446-dim MSE history)
- **FPE signal dim:** 32
- **Exported at:** 2026-04-23T18:45:12.072496+00:00
- **torch / onnx:** 2.8.0 / 1.19.0  (opset 17)
- **Git SHA at export:** `5aceecf48dbf8696a4801a8d73bf5708c04d3b5b`

## Output interpretation (regression)

The FPE graph emits a scalar prediction per time step in
z-scored `edens_ped` units. To recover physical units:

```python
# target_mean=2.580092217753647, target_std=1.606271753311305
y_phys = pred * target_std + target_mean
```

**Validation tolerance note:** RMS tolerance 1e-3 in normalized units ≈ 1.61e-03 (normalized units) for this target (target_std=1.6063).

## Input contract- **MSE history:** `(batch, 50, 446)` — stats per shot.- **MSE mask:** `(batch, 50)` — 1.0 = valid, 0.0 = padding.- **MSE aux:** `(batch, 3)` — `bzn_seconds`, `disrupt_seconds`, `disrupt_coverage`- **FPE sequences:** `(batch, seq_len, 32)` — z-scored signals.- **FPE signal mask:** `(batch, 32)` — per-channel availability.- **FPE padding mask:** `(batch, seq_len)` — 1.0 = valid.See `normalization_params.json` for the exact z-score means andstds used at training time; the per-channel order matches the`fpe_signal_names` list below.<details><summary>FPE signal names (in channel order)</summary>

| idx | signal |
|---|---|
|  0 | `pohm` |
|  1 | `pinj` |
|  2 | `tinj` |
|  3 | `ech_total` |
|  4 | `f1a` |
|  5 | `f2a` |
|  6 | `f3a` |
|  7 | `f4a` |
|  8 | `f5a` |
|  9 | `f6a` |
| 10 | `f7a` |
| 11 | `f8a` |
| 12 | `f9a` |
| 13 | `f1b` |
| 14 | `f2b` |
| 15 | `f3b` |
| 16 | `f4b` |
| 17 | `f5b` |
| 18 | `f6b` |
| 19 | `f7b` |
| 20 | `f8b` |
| 21 | `f9b` |
| 22 | `ecoila` |
| 23 | `ecoilb` |
| 24 | `gasa_cal` |
| 25 | `gasb_cal` |
| 26 | `gasc_cal` |
| 27 | `gasd_cal` |
| 28 | `gase_cal` |
| 29 | `ip` |
| 30 | `ipspr15v` |
| 31 | `bt` |

</details>

## Files

| File | Purpose |
|---|---|
| `mse_encoder.onnx` | Machine-state encoder graph (opset 17) |
| `fpe_encoder.onnx` | Fast-physics encoder graph |
| `model_config.json` | Architecture + task metadata (this card's authoritative source) |
| `provenance.json` | Export-time torch/onnx versions, git SHA, sidecar hashes |
| `normalization_params.json` | Per-channel z-score means + stds for FPE inputs |
| `target_norm.json` | `target_mean` / `target_std` (+ optional `clip_min`/`clip_max`) for de-normalizing regression outputs |

## Validation

Each bundle ships with both random-tensor and real-sample
validation. On the PedestalPredictor GitHub repo, run:

```bash
python -m inference.validate_onnx \
    --model-dir <trial-dir> \
    --onnx-dir onnx_models/edensfit89 \
    --dataset-dir <dataset> \
    --dataset-cls regression \
    --num-samples 10
```

(See [`docs/export_and_publish.md`](https://github.com/SCS-Lab/PedestalPredictor/blob/main/docs/export_and_publish.md) in the GitHub repo for exact per-bundle invocations.)

## License

Licensed under APACHE 2.0 (see repo root).