project-14/soc-scf-trajectories

SOC-SCF Trajectories

Spatially resolved electronic response histories from periodic, spin–orbit-coupled (SOC), noncollinear Quantum ESPRESSO calculations, collected with GPU-accelerated conventional DFT by Project-14.

Release status: The complete SOC-SCF field corpus is mirrored on Hugging Face. All 1,368 raw shards have been verified against their recorded sizes and SHA-256 checksums. See RAW_STATUS.json for the completed inventory.

What is here

Unit	Count
Electronic SCF histories	310
Ordered source evaluations	49,007
Distinct canonical operator/response maps	47,198
Parent roots / geometries / family groups	53 / 163 / 22
Qualified electronic endpoints	32
Backend-converged, not independently qualified	62
Censored histories	216
Unique raw field objects	141,800
Raw field bytes, before lossless packing	1,457,549,308,992 (~1.46 TB)

The dataset pairs spatial charge, magnetization and effective-potential fields with ordered SCF evaluations, calculation settings, scalar diagnostics and endpoint qualification metadata.

An electronic evaluation

At fixed ions and numerical context, the recorded calculation evaluates

input density → effective local potential → spinor Kohn–Sham solve → output density.

Each occurrence has three fields. Density channels are (n, mx, my, mz); potential channels are (V0, Bx, By, Bz). The reader returns arrays with shape (4, nz, ny, nx), C order, x fastest. Density units are electrons/ų; potential units are eV. The B channels are energy coefficients, not Tesla. Native Cartesian spin frames, lattice columns and grids are preserved. A grid point has fractional coordinates (ix/nx, iy/ny, iz/nz); the repeated periodic endpoint is omitted.

The four local potential channels are not the complete SOC Hamiltonian: the nonlocal relativistic pseudopotential/projector terms and numerical context also matter. This is pseudopotential valence-density data, not an all-electron core-density reconstruction. SCF ordering is algorithmic progress, not physical time. Mixing means output density is generally not the next input density.

Files and access

• metadata/trajectories.jsonl: one row per history, with cell/grid, operator settings, qualification, root, family and frozen split.
• metadata/trajectories-{train,validation,test}.jsonl: the same rows partitioned by fixed whole-family split (198 / 43 / 69 histories). These are the dataset viewer configurations, not arrays.
• metadata/occurrences.jsonl.gz: ordered evaluations, source-specific qualification identity, canonical map ID and three field hashes.
• metadata/scalar-histories.jsonl.gz: charge/spin residuals, orbital events and available scalar observations. Preserve nulls; an outer iteration may contain multiple evaluations.
• metadata/density-domain-sidecar.jsonl.gz: available pointwise density diagnostics; missing records remain unassessed.
• metadata/routes/*.json: exact field size/hash/encoding and grid context for each history.
• metadata/inputs/*.in: retained source QE inputs, addressed by their original SHA-256. Numeric filenames are content identities, not execution instructions.
• metadata/fields.jsonl.gz: unique object → raw shard/member lookup. Raw objects retain their original bytes and hashes.
• raw/shard-*.tar.zst: lossless Zstandard-compressed tar shards, each about 1 GiB or less before compression. Members are first-two-hash-characters/full-sha256.
• receipts/shard-*.json: raw shard SHA-256, object count and byte counts after verified upload.
• MANIFEST.json: hashes/sizes for the release metadata and reader.

The first shard contains only one worked endpoint: FeAl2, proto014-feal2-reference130-002, evaluation 96. Read it with:

pip install numpy zstandard huggingface_hub
hf download project-14/soc-scf-trajectories read_example.py --repo-type dataset --local-dir soc-example
python soc-example/read_example.py

Inspect the short reader before executing it. It downloads only the worked shard and its metadata, checks field hashes/header/shape, and verifies 44 electrons per density. Expected shape: (4,72,72,72); charge L1 residual ≈ 3.6909171285e-7 and spin-vector L1 residual ≈ 2.6690981464e-6 electrons/atom.

For a different occurrence, use its route and the field index to fetch only its required shards. Do not recursively download the full repository for an initial analysis. Pin a repository commit for reproducible reads.

Binary encodings

Two original encodings are retained:

1. canonical-f64le: little-endian float64, no header, four contiguous channels in the reader units above.
2. qe-grid1: 64-byte header: ASCII GPOGRID1 followed by seven little-endian signed int64 values (nx, ny, nz, padded_x, padded_y, padded_z, 4). The payload is little-endian float64, four physical-grid channels; its size is 32*nx*ny*nz bytes. Raw potentials are in Ry and densities in electrons/bohr³. Multiply potentials by 13.605693122994 and densities by 1/(0.529177210903**3) to obtain the reader units.

The header/context must match. Check byte counts and original SHA-256 before decoding. No float precision reduction, gauge shift, spin rotation or grid resampling is performed in this release.

Collection and qualification

Instrumented GPU-enabled Quantum ESPRESSO evaluated conventional PBE plane-wave pseudopotential SOC calculations. Recorded contexts use 120/480 Ry wavefunction/density cutoffs and variable native grids. Initial magnetizations, geometries, mixing settings, smearing and selected charge/spin responses vary. Most histories use coarse 2×2×2 k meshes; a few denser controls do not establish physical k-point convergence for the collection. Read each input/context, rather than assuming a shared default.

An admitted intermediate field map is not a converged ground-state label. Qualification is source-occurrence specific; do not transfer endpoint certificates through aliases. Canonical identity binds operator, potential and output density, but excludes input density. Ordered occurrences must therefore remain available for convergence analysis. Independent wavefunction reconstruction applies to selected final maps, not every intermediate field.

Censored histories end without a qualified endpoint and do not prove a system cannot converge. Electronic convergence does not establish the lowest magnetic branch, relaxed ionic geometry, accurate MAE, Curie temperature, coercivity or synthesizability. These are not material-property labels. Full intermediate wavefunctions and mixing matrices are absent; exact calculation replay is not promised.

Appropriate use and limitations

Useful questions include spatial charge/spin residual modes, convergence diagnostics and stopping criteria, field compression, representation studies and conditional operator/residual learning. A successful one-step prediction does not by itself show stable iterative deployment or a DFT speedup.

Evaluations within a history are correlated. Preserve whole-family splits, aliases and response ancestry; never randomly split adjacent evaluations as independent materials. Train, validation and test assignments are fixed at the family level. Coverage is uneven, especially for independently qualified endpoints and opposed/nonparallel initial magnetic configurations. This is a detailed numerical research corpus, not a uniformly balanced survey of magnetic materials or a CPU/GPU speed benchmark.

License and attribution

Scientific data, release metadata and the supplied reader example are licensed under Creative Commons Attribution 4.0 International (CC BY 4.0). Attribute Project-14, SOC-SCF Trajectories (2026) and cite the exact Hugging Face revision. The license does not relicense Quantum ESPRESSO or third-party pseudopotentials; those binaries are not distributed here. GPU collection used compute from SF Compute.

@misc{project14_soc_scf_trajectories_2026,
  author = {{Project-14}},
  title = {SOC-SCF Trajectories},
  year = {2026},
  url = {https://huggingface.co/datasets/project-14/soc-scf-trajectories},
  note = {Cite the exact dataset revision; CC BY 4.0}
}