---
pretty_name: Princeton MagNet Magnetic Material Data
license: cc-by-4.0
task_categories:
- tabular-regression
- time-series-forecasting
tags:
- magnetics
- power-electronics
- material-science
- core-loss
- hysteresis
size_categories:
- 100K<n<1M
configs:
- config_name: steady-state-scalar
  default: true
  data_files:
  - split: full
    path: releases/magnet-steady-state-training-ready/data/scalar/full/*.parquet
- config_name: steady-state-sequence
  data_files:
  - split: full
    path: releases/magnet-steady-state-training-ready/data/sequence/full/*.parquet
- config_name: transient-sequences
  data_files:
  - split: full
    path: releases/magnet-transient-training-ready/data/sequences/full/**/*.parquet
---

# Princeton MagNet data

This is the public, static data library for Princeton MagNet. It contains four
versioned releases in one dataset repository: steady-state raw and
training-ready data, and transient raw and training-ready data. The three
training-ready Parquet views are available as named configurations; the raw
MATLAB and ZIP files are archival downloads. The dataset is not a leaderboard
and contains no challenge scores, participant submissions, or team predictions.

| Release | Repository folder | Measurements | Materials |
| --- | --- | ---: | ---: |
| Steady-state raw | [`releases/magnet-steady-state-raw`](releases/magnet-steady-state-raw/README.md) | 812,492 Combined; 775,865 Cycle | 15 |
| Steady-state training-ready | [`releases/magnet-steady-state-training-ready`](releases/magnet-steady-state-training-ready/README.md) | 775,865 | 15 |
| Transient raw | [`releases/magnet-transient-raw`](releases/magnet-transient-raw/README.md) | 122,399 B/H/temperature sequences | 15 |
| Transient training-ready | [`releases/magnet-transient-training-ready`](releases/magnet-transient-training-ready/README.md) | 122,399 | 15 |

The manifest in each folder gives exact file sizes, SHA-256 checksums, source
lineage, schema version, and validation status. Read the corresponding release
card before using a file; raw and processed releases have different contracts.
All releases currently expose one `full` split rather than an arbitrary
benchmark train/test split.

## Materials

Steady-state: `3C90`, `3C92`, `3C94`, `3C95`, `3E6`, `3F4`, `77`, `78`,
`79`, `ML95S`, `N27`, `N30`, `N49`, `N87`, `T37`.

Transient: `3C90`, `3C92`, `3C94`, `3C95`, `3E6`, `3F4`, `77`, `78`,
`FEC007`, `FEC014`, `N27`, `N30`, `N49`, `N87`, `T37`.

## Use the training-ready data

```python
from datasets import load_dataset

scalar = load_dataset(
    "minjiechen/MagNet", name="steady-state-scalar", split="full"
)
waveforms = load_dataset(
    "minjiechen/MagNet", name="steady-state-sequence", split="full"
)
transient = load_dataset(
    "minjiechen/MagNet", name="transient-sequences", split="full"
)
```

The two steady-state configurations join one-to-one on `measurement_id`.
Steady-state triangular waveforms retain positive and negative duty fractions;
sinusoidal duty is null. The transient release retains variable-length B/H
sequences and a 16 MHz sampling contract. Model users should review quality
flags, operating ranges, and split-leakage guidance in the individual cards.

To download an archival file, use its exact path from the release manifest
with `hf_hub_download(repo_id="minjiechen/MagNet", repo_type="dataset",
filename="releases/<release-id>/<path>")`.

## Provenance and limitations

The selected T37 steady-state Combined file embeds 2023-09-11 while the Cycle
file embeds 2023-08-29. The project owner attests that both derive from the
August 29, 2023 processing and approved publication with this documented
exception. The original MAT bytes retain both embedded dates; independent
Cycle-to-Combined derivation has not been verified. See the steady-state cards
and source catalogs for detail.

The ten transient source ZIPs had contributor names and local paths removed
from Excel package metadata before publication. Scientific worksheet and CSV
content was preserved; published checksums and the sanitization audit are in
the transient raw release. This release is not a claim of physical measurement
accuracy, safety certification, or universal model performance.

## License, citation, and maintainer

All four releases are distributed under [Creative Commons Attribution 4.0
International](https://creativecommons.org/licenses/by/4.0/legalcode)
(CC BY 4.0). Each release includes its own `LICENSE` file.

Please cite Li et al., “How MagNet: Machine Learning Framework for Modeling
Power Magnetic Material Characteristics,” *IEEE Transactions on Power
Electronics*, 2023,
[doi:10.1109/TPEL.2023.3309232](https://doi.org/10.1109/TPEL.2023.3309232).

Maintainer: Princeton Power Electronics Research Lab / Minjie Chen.