README.md

metadata

license: cc-by-4.0
pretty_name: Magnet-Anisotropy-Screening
language:
  - en
size_categories:
  - 1K<n<10K
task_categories:
  - tabular-regression
  - tabular-classification
task_ids: []
tags:
  - chemistry
  - materials-science
  - magnetism
  - permanent-magnets
  - magnetocrystalline-anisotropy
  - DFT
  - crystal-structure
  - rare-earth-free
  - tabular
  - science
configs:
  - config_name: default
    data_files:
      - split: train
        path: magnet_screening.parquet

Magnet-Anisotropy-Screening

A calibrated screening dataset of computed magnetocrystalline anisotropy for uniaxial rare-earth-free inorganic crystals: 3,573 screened structures, 2,242 anisotropy labels (2,044 flagged reliable), with relaxed structures for 88% of labeled entries and a 287-compound high-accuracy calibration tier that quantifies the per-label error.

Magnetocrystalline anisotropy is the property that makes a permanent magnet hard, and it is absent from every large public materials database because computing it requires fully-relativistic density functional theory. Each label here carries the anisotropy constant K1, saturation magnetization Ms, the dimensionless magnetic hardness κ = √(K1/µ₀Ms²), and the easy-axis direction, alongside Curie temperature, hull stability, and, for gate-passing compounds, micromagnetic energy-product estimates over 2,000 sampled microstructures.

Full technical report (17 figures, methods, validation): A Calibrated Screening Dataset of Magnetocrystalline Anisotropy for Rare-Earth-Free Magnet Discovery

At a glance

Rows	3,573 screened structures
Anisotropy labels	2,242 (2,044 κ-reliable)
Relaxed structures	2,787 P1 CIFs, record-linked (88% of labeled rows)
Materials Project provenance	1,958 rows carry mp_id (spot-verified against the live MP API)
Calibration tier	287 compounds recomputed at tightened DFT settings
Format	Apache Parquet (main table, 41 columns) + CSV, JSONL, CIF
Coverage	60 elements, 72 space groups, 3 uniaxial crystal systems
License	CC-BY-4.0

Quick start

from datasets import load_dataset

ds = load_dataset("willgbryan13/magnet-anisotropy-screening", split="train")
row = ds[0]
print(row["record_id"], row["composition"], row["kappa_corrected"], row["easy_axis"])

Or directly with pandas:

import pandas as pd

df = pd.read_parquet("magnet_screening.parquet")
reliable = df[df.kappa_reliable & ~df.k1_outlier.astype(bool)]
hard = reliable[(reliable.kappa_corrected > 1) & (reliable.easy_axis == "001")]
print(len(reliable), "reliable labels;", len(hard), "easy-axis hard compounds")

The error model

The distinguishing feature of this dataset is that its noise floor is measured. A stratified subset of 287 compounds was recomputed at tightened DFT settings (k-spacing 0.16 → 0.10 Å⁻¹, plane-wave cutoff 65 → 80 Ry), and the comparison separates the label error into a bias and a scatter:

• The systematic bias is small and corrected: per-κ-band median corrections of ×1.00 to ×0.96 are applied in the shipped kappa_corrected and K1_corrected_J_per_m3 columns.
• The per-label scatter is the dominant error: in the magnet-relevant bands the interquartile range of the refinement ratio is [0.84, 1.05] (~±16% robust one-sigma), with 48% of labels within ±10% of their refined value.
• In practice, a regression model trained on these labels cannot be validated below a ±15–20% error floor on κ, because the labels themselves move that much under convergence refinement.
• Absolute magnitudes are zero-temperature DFT values, ~1.5–2× room-temperature experiment for hard magnets (the well-documented DFT↔experiment MAE gap).

The calibration pairs ship in data/calibration.jsonl as a held-out gold tier for uncertainty-aware training.

Files

File	Contents
magnet_screening.parquet	main table, 3,573 rows × 41 columns
data/magnet_screening.csv	same table, CSV
data/screening_records.jsonl	complete nested per-record pipeline output
data/relaxed_cifs.tar.gz	2,787 relaxed structures (P1 CIF) + manifest linking record_id → file
data/calibration.jsonl	287-compound high-accuracy calibration tier
data/calibration_summary.json	per-band corrections and scatter statistics
data/METHODS.md	full parameter-level provenance
data/validation_table.md	computed vs literature K1 for canonical hard magnets
summary_stats.json	machine-readable dataset statistics

Schema

Key columns of magnet_screening.parquet (full dictionary with units in data/METHODS.md):

Column	Type	Description
record_id	str	stable identifier, mag-00001 … mag-03573
mp_id	str	Materials Project id of the source structure (1,958 rows; null for off-database variants and ambiguous polymorph pulls)
composition	str	reduced chemical formula
source	str	provenance tag: mp_broad / substitution / template / ordered / generated / doping
verdict, stopped_at	str	cascade outcome and the stage that ended it
actual_space_group, actual_crystal_system	str	symmetry of the relaxed cell
a, b, c, alpha, beta, gamma, volume_A3	float	relaxed lattice parameters (Å, °, ų)
e_above_hull_meV	float	hull distance, meV/atom
Tc_K	float	Curie-temperature estimate, K
Ms_A_per_m	float	saturation magnetization, A/m
K1_J_per_m3, kappa, easy_axis	float, float, str	anisotropy label at production settings; easy axis 001 = axis, 100/010 = plane
kappa_corrected, K1_corrected_J_per_m3	float	per-band convergence-corrected labels, recommended for downstream use
easy_axis_confidence	float	1 − per-band axis↔plane flip rate
kappa_reliable	bool	False where the Ms→0 artifact inflates κ
k1_outlier	bool	extreme-value flag; recompute at refined settings before use
fm_assumption_risk	bool	ferromagnetic-alignment assumption plausibly overestimates Ms
peak_BHmax_kJ_per_m3, peak_Br_T, robustness_score	float	microstructure-sweep ceilings (gate-passing rows only)

Quality control

• Every record relaxed (Orb-v3) before labeling; unphysical cells (nearest-neighbor < 1.5 Å) rejected
• κ artifacts from near-compensated moments flagged (kappa_reliable, 198 labels)
• Extreme single-shot values flagged (k1_outlier, 165 labels)
• Ferromagnetic-alignment risk flagged (fm_assumption_risk, ~50% of rows, consistent with an oxide-heavy set)
• Rejected candidates retained with valid labels as negative examples (easy-plane, soft, metastable)
• mp_id provenance spot-verified against the live Materials Project API (12/12 sampled ids resolve to the same reduced formula)

Distribution snapshots

Sources	mp_broad 2,013 / substitution 1,101 / template 367 / ordered 57 / generated 30 / doping 5
Crystal system (labeled)	trigonal 1,062 / tetragonal 793 / hexagonal 343
Easy axis (reliable)	001 966 : in-plane 1,078 (nearly even)
κ_corrected (reliable, p5/p50/p95)	0.16 / 1.22 / 5.9
Ms (reliable, min/median/max)	0.10 / 0.46 / 1.84 MA/m
Hull distance	67% of reliable labels within 100 meV/atom
Elements	O 825 / Mn 741 / Fe 672 / Cr 414 / Co 272 lead; no lanthanides (uncomputable by construction)

Considerations for using the data

Biases and limitations.

• Uniaxial crystal systems only (tetragonal, hexagonal, trigonal); no cubic anisotropy data.
• Ferromagnetic alignment is assumed throughout; true ground-state orderings are not resolved, only flagged. Magnetization-derived figures are upper bounds where fm_assumption_risk is set.
• Labels are single calculations at screening settings with the quantified per-label uncertainty above; treat this as a screening/ML dataset, not a reference database.
• Curie temperatures and exchange stiffness are method-level estimates and typically run high.
• Rare-earth compounds are absent by construction: the fully-relativistic pseudopotential set behind the anisotropy stage contains no lanthanides. This matches the dataset's purpose (rare-earth-free discovery). Precious-metal compounds are included as high-anisotropy positives.
• Chemistry is Fe/Mn/O-heavy; nickel-rich uniaxial magnets are intrinsically scarce.

Out-of-scope. Cubic magnets, rare-earth magnets, 2D materials, noncollinear or temperature-dependent anisotropy, and experimental coercivity (the microstructure sweep is a proxy model).

Provenance

Screening cascade: Orb-v3 relaxation → symmetry gate → convex-hull stability → Curie temperature and magnetization → TB2J magnetic-interaction extraction over ABACUS DFT (PseudoDojo fully-relativistic norm-conserving pseudopotentials, k-spacing 0.16 Å⁻¹, cutoff 65 Ry; calibration tier at 0.10 Å⁻¹ / 80 Ry) → micromagnetic microstructure sweep (2,000 Latin-hypercube samples per gate-passing compound). Roughly 2,900 anisotropy calculations over an eight-day screening period in June 2026. Property calculations ran as hosted routes on Ouro. Full parameters in data/METHODS.md.

Citation

If you use this dataset, please cite it (a CITATION.cff file is provided alongside this card) and the Materials Project for entries carrying an mp_id.

@misc{bryan2026magnetanisotropy,
  author = {Bryan, Will and Moderwell, Matt},
  title  = {Magnet-Anisotropy-Screening: a calibrated magnetocrystalline-anisotropy
            dataset for rare-earth-free magnet discovery},
  year   = {2026},
  url    = {https://huggingface.co/datasets/willgbryan13/magnet-anisotropy-screening},
  note   = {ghost-projects, Project 014. v1.0}
}

Companion work: GPSK-300 (structure generation) and the GPSK-Inorganic-Crystals corpus.

License

Released under CC-BY-4.0. Source structures from the Materials Project retain their identifiers in the mp_id column; see LICENSE for attribution terms.