SOURCES_AND_PROCESSING.md

# Neo Dataset Sources and Processing Declaration

## Scope of this declaration

Neo is a derived, mixed-granularity atomistic training corpus. It aggregates
published periodic materials, molecular response, finite-field water, Born
effective charge, and magnetic spin data into one mask-aware HDF5 schema. Neo
does not claim authorship of upstream first-principles calculations.

The conversion pipeline does not infer or fabricate missing physical labels.
It copies documented upstream quantities, performs the unit conversions listed
below, assigns provenance and group identifiers, and explicitly masks labels
that are absent or physically incompatible with a shared loss.

The machine-readable companion to this statement is
`provenance/source_registry.json`. File integrity records are in
`checksums.sha256`, `manifest.json`, and the validation reports under
`reports/`.

## Upstream source register

| Neo source IDs | Upstream record | Upstream terms | Redistribution status |
| --- | --- | --- | --- |
| `OMat24/*` | OMat24 (Open Materials 2024), official FAIR Chemistry data page; manuscript DOI `10.48550/arXiv.2410.12771`; ColabFit subdataset records | CC BY 4.0 | Cleared with attribution and modification notice |
| `MPtrj-v2022.9-large`, `MPtrj-v2022.9-magnetic`, `MPtrj-static-recovered` | MPtrj, DOI `10.6084/m9.figshare.23713842.v2` | MIT | Cleared with attribution |
| `JARVIS-DFT-3D-v2025.09`, `JARVIS-DFT-DFPT-v2025.09` | JARVIS-DFT 3D, DOI `10.6084/m9.figshare.6815699.v11` | CC BY 4.0 | Cleared with attribution |
| `QM7-X` | QM7-X, DOI `10.5281/zenodo.4288677`; article DOI `10.1038/s41597-021-00812-2` | CC BY 4.0 | Cleared with attribution |
| `DES15K`, `TorsionNet500`, `SPICE_dipeptides` | SO3LR data release, DOI `10.5281/zenodo.14779793` | CC BY 4.0 | Cleared with attribution |
| `SCFNN/zero-field`, `SCFNN/finite-field` | SCFNN training data, DOI `10.5281/zenodo.5521328` | CC BY 4.0 | Cleared with attribution |
| `DeepSPIN-NiO` | `yangtengleo/DeepSPIN` commit `526ade353906f21cf4d8cb32db3d53ce83e1ed53` | GPL-3.0 | Redistributable subject to GPL-3.0 notices and source access |
| `BEC/H2O`, `BEC/MAPbI3`, `BEC/dimer` | `MLFF_and_BEC` archive associated with Schmiedmayer and Kresse, article DOI `10.1063/5.0217243` | Article: CC BY 4.0; archive: no standalone license found | **Archive-level rights review remains open** |

The article license for the last row is evidence about the publication, not by
itself a conclusive license for a separately supplied data archive. The three
`BEC/*` source IDs are therefore not treated as cleared. Every current portable
tier includes some of these rows.

### OMat24 chain of custody

The OMat24 portions of SE, Plus, and Max originate from the eleven OMat24 training
subdatasets mirrored by ColabFit and documented by FAIR Chemistry. Their source
cards identify Luis Barroso-Luque, Muhammed Shuaibi, Xiang Fu, Brandon M. Wood,
Misko Dzamba, Meng Gao, Ammar Rizvi, C. Lawrence Zitnick, and Zachary W. Ulissi
as dataset authors, link the OMat24 manuscript at
<https://doi.org/10.48550/arXiv.2410.12771>, and declare CC BY 4.0. Neo removes
duplicate `configuration_id` records, assigns deterministic leakage-safe
splits, selects rows for each Composite tier, and repackages retained Arrow records
inside HDF5. It does not alter the retained float64 geometry, energy, force, or
stress values. `selection/source_row_index`, source paths, and checksums retain
the audit trail. These are modified and repackaged derivatives, not official
Meta, FAIR Chemistry, or ColabFit releases.

Meta Platforms, Inc., FAIR Chemistry, ColabFit, and the named OMat24 authors
have not endorsed Neo. Their names and marks are used only for provenance and
attribution. Users must comply with CC BY 4.0, cite the authoritative OMat24
record, preserve attribution and the license link, and identify Neo's
selection, deduplication, split assignment, and repackaging changes.

## Chain of custody and integrity

The following identifiers anchor the data used by the conversion. Local paths
are deliberately omitted from public metadata.

| Artifact | Bytes | Integrity identifier |
| --- | ---: | --- |
| Official MPtrj JSON | 12,188,168,685 | MD5 `50ead5f27f9a4f6beb7564c4188f1e9f`; SHA-256 `05a675acd4f17a1dadd8a4ea6ea79fe58d3ffb28ca20277383b8f990d4787cf6` |
| Official JARVIS 3D ZIP | 48,447,610 | MD5 `c3179161ef0d029cf1fa70da463aac6e`; SHA-256 `f9e0a3309f0000d5de1ec9e49c93963109ea45e63f451103f8f6595d2eabf7f5` |
| Official QM7-X `8000.xz` | 89,426,884 | MD5 `c893ae88b8f5c32541c3f024fc1daa45`; SHA-256 `9868f7781bf7f0a53c24d982b5dc31aa11d059d220d1894d1beaba9671526d73` |
| Decompressed QM7-X HDF5 | 826,408,960 | SHA-256 `04c4ed110eae21f719bc361e4a550d9b628232063df844e75d559a69e070679a` |
| SO3LR DES15K HDF5 | 106,850,728 | SHA-256 `a0cc8b509ae3d10d3612b47ae973a4a28b6b661424b36c816b0371ac15c9f794` |
| SO3LR TorsionNet500 HDF5 | 108,233,896 | SHA-256 `2b4cfead21ee9f3b7fdc92e9d85697098c1aca267ec6170e50d0f29177648016` |
| SO3LR SPICE dipeptides HDF5 | 399,683,424 | SHA-256 `f779afd3fb466c011cd21a22bb7f21c1c861ffd6caee9db57439feaf65577e35` |
| Official SCFNN `train_test_data.tar.gz` | 178,468,151 | MD5 `5361bc9ceb59fdd42250ee2ff1f06a8d` |
| DeepSPIN source | Git commit | `526ade353906f21cf4d8cb32db3d53ce83e1ed53` |
| `MLFF_and_BEC.tar.gz` supplied source | 81,959,223 | SHA-256 `957cb5dd80918a0ccc62bbb4a5a2f32ad5b433933df3c370de79e18691b9c98a` |
| Combined `static.extxyz.gz` | local aggregate | SHA-256 `40569a015be62218e34cba63623317819e1ec777a082a3fc3678a1c2ab797484` |
| Combined `response.extxyz.gz` | local aggregate | SHA-256 `06d6245c892558d69d50562e909ba766e798d29449fac3e8dd911196737161db` |

The combined extXYZ archives are intermediate aggregation artifacts and are not
the preferred public provenance record where an official source is available.
Their fixed half-open frame ranges are:

| Archive range | Contents | Canonical treatment |
| --- | --- | --- |
| `static[0:200000]` | MPtrj structures | IDs and MP2020-corrected energies recovered against the source data; a balanced 40,000-material shard was retained |
| `static[200000:201590]` | SCFNN zero-field water | Recovered into `neo_l2_scfnn_field.h5` |
| `response[0:101993]` | Legacy molecular response prefix | Replaced for canonical construction by the official QM7-X and SO3LR HDF5 sources |
| `response[101993:105179]` | SCFNN finite-field water | Recovered and geometry-grouped with zero-field variants |
| `response[105179:105729]` | H2O, MAPbI3, and dimer BEC records | Recovered into `neo_l2_bec.h5`; rights review remains open |

## Source-specific transformations

### MPtrj

- Structures, cells, forces, VASP stress, and atom-wise `magmom` values are
  read from the official archive.
- Total energy is `corrected_total_energy`, which applies Materials Project
  MP2020 compatibility to GGA/GGA+U calculations.
- A site with `abs(magmom) >= 0.05 mu_B` receives a z-directed unit spin with
  the sign of `magmom`; its vector magnetic moment retains the signed magnitude.
  Smaller values become zero-valued nonmagnetic sites inside a structure whose
  spin mask is active. Missing structure-level `magmom` arrays are masked, not
  replaced by zeros.
- The large core keeps all required legacy sample IDs and at most four frames
  per material family. All frames sharing an `mp_id` share one split.
- The standard corpus uses one selected magnetic frame per material and a
  separately recovered static subset, then removes poorer duplicates by
  material group.
- Source stress is the VASP compressive-positive tensor in kBar. Neo
  symmetrizes it and multiplies by `-0.0006241509074460763` to store
  tensile-positive Cauchy stress in eV/angstrom^3. No volume division is
  applied because the source value is already intensive.
- The complete archive was streamed once to match canonical sample IDs. Stress
  was recovered for 40,000 static records, 24,000 magnetic records, and
  505,736 Large-core records; 72,929 Large-core stress records also carry spin
  labels. MPtrj does not provide the paired reference spin state required for
  a differential magnetoelastic target.

### QM7-X

- Geometry and the published `ePBE0`, `eMBD`, `ePBE0+MBD`, force, Hirshfeld,
  dipole, polarizability, and dispersion fields are read directly from the
  official HDF5 shard.
- `energy = ePBE0+MBD`, `energy_base = ePBE0`, and
  `energy_dispersion = eMBD` in the source-specific canonical shard.
- `forces`, `forces_base`, and `forces_dispersion` map to `totFOR`, `pbe0FOR`,
  and `vdwFOR`.
- Atomic dipoles are multiplied by `0.529177210903` to convert `e*bohr` to
  `e*angstrom`.
- Molecular and atomic polarizabilities are multiplied by
  `0.529177210903^3` to convert `bohr^3` to `angstrom^3`.
- Atomic C6 values are multiplied by
  `27.211386245988 * 0.529177210903^6` to convert
  `Hartree*bohr^6` to `eV*angstrom^6`.
- All conformers of a QM7-X molecule share a group and split.
- QM7-X absolute energies are masked from the default shared periodic MPtrj
  energy loss. Response labels remain active.

### SO3LR molecular families

- DES15K, TorsionNet500, and SPICE dipeptide records come from the official
  SO3LR data release and its PBE0+MBD/tight calculations.
- Atomic dipoles use the same `e*bohr` to `e*angstrom` conversion as QM7-X.
- Total charge is the nearest integer to the sum of published Hirshfeld charges.
  The original atomic charges are retained.
- Parent identities are derived from the documented family naming convention,
  so related conformers and geometries remain in one split.
- Their absolute energies are retained in independent shards but masked from
  the default mixed shared-energy loss.

### JARVIS-DFT

- The L1 shard converts published OptB88vdW energy per atom to total energy.
- The DFPT shard reads each published per-ion `born_charges` 3x3 tensor from
  `vasprun.xml` without projecting it onto the acoustic sum rule.
- The same raw archives provide a VASP stress tensor and the electronic
  `PIEZOELECTRIC TENSOR for field` block. Neo stores stress with the same
  tensile-positive conversion used for MPtrj and expands VASP columns
  `XX YY ZZ XY YZ ZX` into a full tensor with no engineering-shear rescaling.
- Of the selected 120 archives, 112 pass the BEC sanity gate and jointly carry
  stress, BEC, and electronic clamped-ion piezoelectric labels. They form the
  explicit L2 electromechanical bridge in every Tiny-Large tier.
- Records are rejected if `max(abs(BEC)) > 50 e` or the maximum component of
  the acoustic-sum residual exceeds `0.5 e`. Rejections are recorded in the
  build metadata.
- Structure-total magnetization and dimensionless dielectric constants are not
  relabeled as atom-wise moments or molecular polarizabilities.

### SCFNN

- The official dataset contains periodic 64-water boxes at zero and finite
  electric field. The Neo conversion retains supplied energy, forces, field,
  and dipole values in its independent shard and sets the known neutral total
  charge to zero.
- A SHA-256 hash of atomic numbers, positions, cell, and PBC rounded to `1e-7`
  groups field variants of the same geometry. Grouped variants cannot cross
  train, validation, or test boundaries.
- The default mixed corpus keeps field, dipole, and total-charge supervision;
  its SCFNN absolute energy and force reference is not mixed into MPtrj.

### DeepSPIN NiO

- The 48-site raw representation contains 32 real atoms plus 16 pseudo-atoms.
  The periodic minimum-image displacement from each Ni atom to its pseudo-atom
  is divided by the documented virtual length (`0.4 angstrom`) to recover the
  unit spin direction.
- Magnetic moments use the documented magnitude `1.2737 mu_B`.
- Effective spin field is `0.4 * pseudo_atom_force`, in `eV/spin`.
- The author's five 20-frame blocks are preserved: blocks 0-3 are train and
  block 4 is validation.
- DeltaSpin energy and real-atom forces remain in the independent shard but are
  masked from the default mixed energy/force loss.

### Schmiedmayer-Kresse BEC archive

- The supplied archive contains 100 liquid-water, 300 MAPbI3, and 150 water
  dimer DFPT structures. This matches the systems and calculations described
  in DOI `10.1063/5.0217243`; the tar owner is `bernhard`, and the included VASP
  settings match the article methodology.
- Neo extracts atomic numbers, positions, cell, forces, energy, and the
  published per-atom 3x3 Born effective charge tensors. The mixed corpus keeps
  only BEC, zero field, and known neutral total charge for these records.
- The raw OUTCAR files include VASP/POTCAR notices and must never be uploaded.
- Identification, scientific attribution, and current public hosting do not
  resolve the absent archive license. Written confirmation or removal of these
  records is still needed for a fully cleared redistribution record.

## Aggregation, masking, and energy domains

`config/neo_mixed_source_policy.json` and
`config/neo_large_source_policy.json` define source order, retained labels,
deduplication keys, and per-group limits. The conversion writes every supported
label dataset densely, but `masks/<label>[i]` is the sole authority for whether
structure `i` owns that target.

The default shared energy/force domain is MPtrj with MP2020-compatible energy.
QM7-X, SO3LR, SCFNN, JARVIS OptB88vdW, local BEC, and DeepSPIN energies have
different electronic-structure references. They are either kept in independent
source shards or masked in mixed tiers. No offset fitting is used to pretend
that these absolute references are interchangeable.

`J_effective`, `Di`, `Di_effective`, and `DMI_effective` are absent from the
released tiers. Their masks are all false. Zero-valued placeholders outside a
mask are not physical labels.

## Group-safe splits

Most groups use a stable SHA-256 bucket: 80 percent train, 10 percent
validation, and 10 percent test. The grouping unit is the physical parent, not
the individual row: MPtrj material, QM7-X molecule, SO3LR parent, SCFNN
geometry, or DeepSPIN author block. Consequently, related trajectories,
conformers, and field variants cannot leak across splits.

The portable-tier builder also guarantees that every element appearing in an
evaluation split appears in train. This is a coverage rule, not evidence of
predictive accuracy.

## Portable tier sampling

Tiny and small are deterministic nested subsets of standard:

```text
tiny subset small subset standard
```

Sampling seed `20260720` uses source weights proportional to `sqrt(N)`, then
stratifies over source, fixed split, active-label family, chemical complexity,
atom-count class, and element. Tiny uses a non-rare maximum of two structures
per group; small uses four. Sources containing at most 128 structures are
preserved in full. Missing labels and their masks are copied unchanged.

The exact configuration and audit are in `config/neo_dataset_tiers.json` and
`reports/neo_tier_hierarchy_audit.json`.

## Validation performed

- Canonical schema and pointer consistency.
- Finite geometry and finite values under every active label mask.
- Unique sample IDs and no group crossing split boundaries.
- No response family crossing split boundaries.
- Charge-sum residual, active-spin unit norm, and BEC acoustic-sum diagnostics.
- Fully 3D periodic nonsingular cells for stress, finite symmetric stress, and
  explicit source method/sign/unit metadata.
- Symmetry of the two strain axes in piezoelectric and magnetoelastic tensors.
- Mechanism-stratified counts for total stress, stress+BEC+piezoelectric,
  stress+spins, and paired magnetoelastic response.
- Exact nesting of tiny in small and small in standard.
- Source and active-label coverage in every portable tier.
- Train coverage of all elements appearing in validation or test.
- Full-physics forward/backward smoke tests on CPU and Apple MPS.

The current paired magnetoelastic count is zero. This is a scientific negative
result, not missing bookkeeping: total magnetic stress cannot be converted into
`stress(S) - stress(S_ref)` without a same-geometry, same-method reference
calculation. The released presets therefore keep `w_magnetoelastic = 0`.

These checks establish structural validity, provenance traceability, numerical
sanity, and leakage controls. They do not establish production accuracy. That
requires a converged model and held-out, domain-specific evaluation.

## OMat24 Plus curriculum

Plus combines a deterministic material-family quarter of all 11 locally
verified OMat24 corpora with the complete Neo Large corpus. A material is
selected when the first byte of
`SHA-256("neo-plus-v1|" + material_id)` is below 64. Exact duplicate
`configuration_id` values are removed before the final selector is written;
all configurations from one material share a split determined by the
`neo-split-v1` namespace.

The portable SE HDF5 embeds every Standard geometry, label, mask, and provenance
field and selects an exact stratified `1/180` of the Max OMat24 selector. Plus
and Max embed every Large geometry, label, mask, and provenance field. All
three convert selected rows from the 635 OMat24 source shards into ragged
packed HDF5 arrays, retaining source names, original row indices, byte
checksums, and float64 numerical values without quantization. Unselected rows
are not embedded. The runtime reads contiguous packed batch spans, so each
portable artifact has no dependency on the original OMat24 directory. A final
portable SE, Plus, or Max artifact must declare both
`sources/neo_large.embedded = true` and
`sources/omat24.embedded = true`; a file containing only paths to either source
is an intermediate manifest and must not be published as a portable tier.

The training curriculum keeps incompatible absolute-energy references apart.
Base fits OMat24 PBE+U atomic references from OMat24 only. Response uses Large
records carrying L2/L3 response or spin labels. Mixed Joint balances the
embedded Large response corpus against rotating OMat24 foundation windows;
Large absolute energy is masked in that stage while compatible force and
response labels remain active.

## OMat24 Max curriculum

Max scans every configuration row in the same 11 verified OMat24 corpora.
Rows are deduplicated globally by exact upstream `configuration_id`, while
material-family IDs continue to define leakage-safe train, validation, and
test assignments. The current build scanned 100,824,534 source rows, removed
154,252 repeated IDs, and retained 100,670,282 unique OMat24 configurations.
It then adds the complete 613,267-structure Large corpus, for 101,283,549
training records in total. Max uses the same Base, Response, and Joint energy
reference policy as Plus.

## Reproducibility boundary

All executable conversion, sampling, validation, and release-staging logic is
in `Datasets_Preparation.py`. Runtime HDF5/Parquet readers remain in
`E3_miu_GNN.py`. The public Hugging Face staging package omits
raw archives, selection-candidate dumps, local paths, checkpoints, and VASP
outputs. The original raw archives must be obtained from their authoritative
sources under their respective terms.