Published records
Datasets exported from this repository and deposited in PSDI Data Collections, where they have a permanent identifier and can be cited.
These are snapshots. The AiiDA database remains the authoritative calculation record; a published dataset is a documented view of it at one point in time.
Quantum ESPRESSO no-spin SCF k-point convergence
Two records, one body of calculations. Same structures, same settings, same
pw.x runs — they differ in what was extracted. See
QE SCF k-point sweeps for the settings and how to
run them yourself.
52713-55d86 — the converged mesh per structure
52713-55d86 · v2.0 ·
CC BY 4.0
The converged k-point mesh for 17,757 MC3D structures, numbered on the 1-based ladder (rung 1 the Γ-only (1, 1, 1)
mesh) and built with the resolution floor min_k_distance = 0.03 Å⁻¹ rather
than a per-axis k-point cap. No spin polarisation, SSSP PBEsol pseudopotentials,
every mesh unshifted and therefore gamma-inclusive.
Convergence is the first of three consecutive ladder meshes whose total energies agree within 1 meV per atom. Energy only — no force criterion.
| File | Contents |
|---|---|
convergence_summary.csv |
17,757 rows: source_db_id, k_index, k_dist_interval, k_mesh |
CIF_files.tar.gz |
18,220 structures, CIF_files/<source_db_id>.cif |
The archive carries more structures than the table has rows: 463 structures were calculated but never met the criterion within the range of meshes swept, so they have a structure file and no converged answer.
k_index needs both its base and its floor to mean anything; the record's
README.md and manifest.json carry both, and a consumer reads them from there
rather than assuming.
See convergence criteria for how labels are assigned,
and the record's own README.md for the full definition and reproduction code.
mcpnq-g1j55 — the full per-calculation dump
mcpnq-g1j55 ·
CC BY 4.0
The complete DFT data behind Automatic generation of input files with optimised k-point meshes for Quantum ESPRESSO self-consistent field total energy calculations — the training set for the paper's machine-learning models. The same calculations as the record above, expressed as a k-distance, and a raw per-calculation dump rather than a convergence-label table. Take this one when you want the QE outputs themselves: cutoffs, Fermi level, symmetry counts, wall time.
| File | Contents |
|---|---|
data.tar.gz |
Unpacks to the two entries below |
summary.csv |
Per-material k-point convergence: Goldilocks-optimised mesh, MC3D reference mesh, k-distance metrics, and medium / well / ultra levels |
structure_calc_details/ |
One directory per structure: the .cif plus the full QE output (energies, cutoffs, Fermi level, symmetry counts, wall time, convergence notes) |
This record predates the k-mesh ladder convention and carries meshes and
k-distances directly, not a k_index.
Quantum ESPRESSO nscf band structures
r3byg-xp284 · v1 ·
CC BY 4.0
See QE nscf band structures for the settings and how to run them yourself.
Non-self-consistent band-structure calculations for 19,405 MC3D structures: one summary row per material, the primitive cell the bands were computed on, and the eigenvalues along the high-symmetry k-point path. Every structure has all three — there is no row without a structure file and none without a band table.
Same settings and the same family of structures as the SCF records above. Those answer which mesh is dense enough; this one answers what the band structure says about the material.
| File | Contents |
|---|---|
nscf_band_summary.csv |
19,405 rows: identifiers, Fermi energy, metallicity, band_gap_ev |
CIF_files.tar.gz |
19,405 primitive cells, CIF_files/<source_db_id>.cif |
bands.tar.gz |
19,405 eigenvalue tables, bands/<source_db_id>.csv |
example.py |
Loads all three from the tarballs and plots one band structure |
9,854 rows are metal and 9,551 insulator. band_gap_ev is populated for
exactly the insulators and empty for exactly the metals.
metallicity is a zero threshold, not a physical one
A row is insulator when the band structure has a gap at the Fermi level, with
no tolerance at all: the smallest gap in the table is 0.0016 eV, and nothing
was rounded down. A PBEsol gap of a few meV is well inside the error of the
method, and a practitioner would treat such a system as metallic.
No thresholded label is published. A threshold frozen into an immutable record
becomes a convention every consumer then has to discover and match — the same
failure that forced the first dataset's k_index column to be recomputed
wholesale. The gap is the fact; the cut belongs to the person using it, and
band_gap_ev is published raw so it can be made.
Anything downstream that trains on this column should name the threshold in its own target name, so that a later dataset built on a different cut cannot be mistaken for this one.
The k-point coordinates were reconstructed, and checked
The eigenvalue tables as calculated carried no k-point coordinates. The path
came from SeeKpath through the standard PwBandsWorkChain, so it was
regenerated per structure — the calculations did not use a single sampling
density,
so the density was determined structure by structure.
A reconstruction was accepted only when it matched the calculation's own record of the labelled points on all four of: number of k-points, labels, label positions in the table, and label coordinates. That makes the match a verification rather than an assumption, and it doubles as a detector.
Two groups are excluded from the record entirely — structure, summary row and bands alike:
| Excluded | Why |
|---|---|
| 104 structures | the regenerated path did not match, rather than publish coordinates that might not be theirs |
| 80 structures | their eigenvalue tables were not in the working archive |
The eigenvalues themselves are copied unchanged from the calculation output.
Publishing another one
See Publish a dataset.