kazeevn/LeMat-Bulk-fixed

LeMat-Bulk, compatible_pbe, with stale Materials Project energies fixed

This is LeMaterial/LeMat-Bulk compatible_pbe at revision 0dc17eea904b860ad7288141e9870f67f8e6bb2c, with the Materials Project (MP) rows re-checked against MP's current task database, a few broken records removed, and the forces and stress that upstream left empty filled in. The schema, column order and shard layout are unchanged. Every row not listed in the audit is identical to upstream. Only compatible_pbe is provided.

	rows
upstream compatible_pbe	5,335,299
DFT record replaced	17,998
removed, each with a stated cause	131
kept but flagged	832
forces and stress added (otherwise unchanged)	27,318
this dataset	5,335,168

The audit config has one row per replaced, removed, flagged or force-filled entry. It gives the old and new energies, the MP task ids and a cause column in plain text.

No row has empty forces or stress_tensor. Upstream had 30,679 MP rows with both empty, because LeMaterial read MP's empty output.forces. Here they are filled from MP's S3 task documents (s3://materialsproject-parsed):

• Rows whose DFT record is otherwise upstream's (27,332, of which 14 are also flagged). The values come from the last ionic step of the same MP task. That task was identified by matching the upstream energy and geometry exactly; on 500 control rows the procedure reproduces upstream forces and stress bit-exactly.
• Replaced rows whose new MP task has empty output fields (2,512). The values come from that task's last ionic step, used only where the step's geometry (to 1e-4 Å) and energy (to 1e-5 eV) are the task's output. All 2,512 matched exactly.

Stress is in kBar with VASP's sign convention, as upstream. The audit column forces_source says which source each row's forces came from.

The problem

LeMat-Bulk's MP rows carry the final energy of one MP task. MP's task database shows three things about them:

• Deprecated tasks. 15,002 rows use a task MP has since deprecated:
  • 653 are plain GGA where MP's GGA/GGA+U mixing scheme requires GGA+U, or the reverse. These are oxides and fluorides of V, Cr, Mn, Fe, Co, Ni, Mo and W. LeMat-Bulk was designed to prevent this kind of mixing.
  • 2,463 used POTCARs that differ from MP's current calculation, such as Mg_pv 13Apr2007, Ge_d 03Jul2007, Li_sv 10Sep2004 and Ce 23Dec2003.
  • 263 have U values that differ from MP's, for example U on O instead of the metal.
  • 1,343 come from MP's retired batch vol_deprecated_mar_2019. Its mp-1079094 (C2/m Sb₂Te₃) reports −4.681 eV/atom; a static run on the same structure gives −3.201.
• Untraceable energies. 3,744 rows carry an energy that matches no task MP still serves.
• Merged ids. 100 ids are no longer MP materials, because MP has merged them into other materials as tasks. They are followed to their parent material.

A too-low energy becomes a vertex of the convex hull and pushes up e_above_hull for everything else in its chemical system. Before this fix, the experimental ground state of Sb₂Te₃ (mp-1201) sat 1.018 eV/atom above a raw-energy PBE hull built from compatible_pbe.

LeMat-Bulk discussion #6

The three entries in discussion #6:

entry	upstream (eV/atom)	here	MLIP median	what happened
mp-1400697 Bi₂O₈Pd₃Pt	−42.861	−1.926	−2.294	untraceable energy; replaced by MP's static on the same structure
mp-725221 CN₄Cl	−0.611	−6.464	−6.471	untraceable energy; replaced by MP's current calculation
mp-554674 Ga₃N₃OZn	−5.707	−5.707 (kept)	−5.628	upstream is right; MP's current value is the broken one (below)

Why MP's current mp-554674 energy is wrong. MP's GGA_GGA+U thermo document uses the 2021 static mp-1840627, at −0.356 eV/atom. That static, and the 2020 statics mp-1347023 and mp-1771371, all ran on a corrupted input structure:

• It has the same lattice and the same fractional coordinates as the 2014 static mp-915025, except that one fractional coordinate is negated.
• In this γ = 120° cell that is not a symmetry operation or a mirror; it shears the crystal.
• Negating that coordinate reproduces MP's input exactly, including a 1.21 Å Ga–O contact (real Ga–O bonds are 1.9–2.0 Å).

POTCARs, electron count and settings are otherwise the same. The 2014 relaxation (LeMat's source) and the 2014 static agree with each other at −5.708 eV/atom, with a shortest contact of 1.89 Å.

Method

All energies are raw, uncorrected VASP energies. No MP2020 corrections are applied, and no MP formation or hull energies are used.

1. Provenance. For every MP row, the task whose output energy reproduces the LeMat-Bulk energy (to 1 meV) was identified among the material's GGA and GGA+U tasks. It was compared with the task MP's GGA_GGA+U thermo document uses now. The data came from the MP API on 2026-10-14 (database 2026.04.13).
2. Candidates. Full task documents were read from MP's public S3 bucket materialsproject-parsed. Each candidate's final geometry was compared with the LeMat-Bulk structure: lattice and permuted sites to 1e-3 Å count as "same", to 0.05 Å as "near". Only tasks MP has not deprecated, with MP's run type for the material, were considered.
3. Choosing between DFT values. Where valid DFT calculations disagree, the MLIP energies of LeMat-Bulk-MLIP-Hull-All choose between them. These are ORB-v3 conservative and direct, UMA, MACE-MP-0b3 and MACE-OMAT-0, evaluated on the upstream structures. On clean rows the MLIP median lies within 0.2 eV/atom of DFT 99.5% of the time. The MLIPs only choose between DFT calculations; no row is removed because it disagrees with them.
4. Rules.
   • Stale rows (deprecated task, wrong run type, or untraceable energy) take MP's current task when it is on the same or a near structure. Otherwise they take another live same-structure task. Otherwise MP's current task replaces the whole DFT record: structure, energy, forces, stress and magnetic moments. A candidate must lie within max(0.2, |upstream − MLIP| + 0.05) eV/atom of the MLIP median.
   • Live rows (a valid task, but not the one MP uses) are kept, unless upstream is more than 0.2 eV/atom off the MLIPs and MP's task is within 0.2 and 0.05 closer. This switches 191 rows.
   • Same-structure candidates are always preferred, so that the MLIP energies in LeMat-Bulk-MLIP-Hull-All still describe the row.

Field mapping was checked against 600 control rows, where the chosen task is the upstream one. Energy, forces, stress, site magnetic moments and positions are bit-exact. total_magnetization is the sum of site moments and last_modified is the MP task's last_updated, both as in upstream. No replacement structure has an unphysical contact: the worst shortens the shortest bond by 12%.

Removed rows (131), each with a cause

The cause column of the audit has the full evidence for each row.

• Non-standard POTCARs with no valid replacement (112 MP rows). These rows use Mg_pv 13Apr2007 or Ge_d 03Jul2007. MP has deprecated the materials and has no other valid calculation for them. Measured against the standard POTCAR on same-structure pairs, the offsets are +0.097 eV per Mg atom (116 pairs, IQR ±0.004) and +0.107 eV per Ge atom (22 pairs).
• Spurious Cr-centred electron population (10 Alexandria Cr fluorides/oxides).
  • Alexandria's own Bader charges put 12.6–22.9 electrons on Cr. In 24 normal Alexandria Cr fluorides/oxides the range is 10.3–11.7, and the anions here are normal.
  • The energy deficit, 25–54 eV per cell, is 4.0–4.8 eV per excess electron in 8 of the rows (5.2–6.4 in the two mixed-metal ones). The excess comes in quanta of about 6 electrons per Cr.
  • Within single relaxations the energy jumps by the same quanta. One run without the excess (CrLi₂₄Mn₁₁O₃₆, first relaxation) agrees with MP's identical structure.
  • A variational PBE+U calculation with a physical PAW setup cannot do this. It is consistent with a PAW ghost state.
  • Alexandria's published records carry the same energies.
• Overlapping atoms (3 rows). agm001439530 has Bi–Cu at 1.09 Å; oqmd-7966826 and oqmd-5708135 (duplicates) have Fe–Mg at 0.90 Å. That is 0.3–0.4× the sum of atomic radii, so the PAW spheres overlap.
• Coincident atoms (1 row). oqmd-2969647 (Cu₅Mg₈) stores 13 atoms on 8 positions, with each Cu exactly on a Mg, which no calculation can have used. Its nsites says 8, and its forces and magnetic moments are all zero. OQMD has no Cu₅Mg₈ entry, and its only Cu₅Mg₃ entry is a different structure.
• Energy belongs to another structure (2 OQMD rows).
  • oqmd-1922538 has a total energy bit-identical to oqmd-2011591, whose cell is 43% smaller.
  • oqmd-2002394 has the same total energy as oqmd-2826002, a different CoC₃ structure.
• Not reproducible (2 OQMD rows).
  • oqmd-3005551: the identical crystal is computed by MP, Alexandria and three more OQMD entries, all 2.13 eV/atom higher.
  • oqmd-3245677: three other OQMD calculations of the identical structure are 0.73 eV/atom higher.
• Collapsed relaxation (1 MP row). mp-1518724's only task went through a 40 ų volume collapse (−20,101 kbar) in 4 of 99 steps. Its final energy is 3.45 eV/atom below step 1 at a volume within 5%, with −0.4 μB on every O. Alexandria's identical structure is 3.40 eV/atom higher.

Kept but flagged (832)

• MP-deprecated materials with no stated reason (793). MP returns no deprecation_reasons. These rows use standard POTCARs and MP2020 U values, and agree with the MLIPs.
• Stale MP tasks with no valid replacement (35). These agree with the MLIPs.
• Unexplained MLIP gap (4). oqmd-3911617, oqmd-5269036, oqmd-2982764 and agm001685190 sit more than 0.5 eV/atom below all five MLIPs, but no calculation-level defect was found.

Result

Rows whose DFT energy is below all five MLIPs, excluding rows whose structure changed:

source	> 0.2 eV/atom before → after	> 0.5 eV/atom before → after
MP	268 → 15	115 → 0
Alexandria	5,041 → 5,041	5 → 5
OQMD	264 → 264	3 → 3

The remaining rows at 0.2–0.5 eV/atom are dominated by Pu, U, Np, Pa and Tc systems, where the MLIPs are the weak side. The remaining rows above 0.5 eV/atom are 4 Pu Heuslers and the 4 unexplained rows above.

Raw-energy PBE e_above_hull (eV/atom), upstream → this dataset:

entry	before	after
mp-1201 Sb₂Te₃ R-3m (ground state)	1.018	0.000
mp-1079094 Sb₂Te₃ C2/m	0.000	0.462
mp-1182783 Bi₂Te₃ C2/m	0.000	0.634
mp-28305 K–Mo–O	0.000	0.389
mp-725221 CN₄Cl	6.828	0.975

Ytterbium

• All Yb data here is on the Yb_2 scale. There are no MP rows containing Yb, upstream or here. All 10,032 Yb rows come from Alexandria and are unchanged. On 1,210 of them whose structure MP also has, Alexandria's energy equals MP's Yb_2 energy: the median difference is +0.5 meV/atom, or 0.002 eV per Yb atom. All 8,274 of MP's GGA/GGA+U Yb tasks use PAW_PBE Yb_2.
• MLIPs. MACE-MP-0b3 (MPtrj) matches these rows: 0.00 eV per Yb atom. ORB-v3, UMA and MACE-OMAT were trained on OMat24, which used the PBE_54 Yb_3 POTCAR (OMat24 record). They sit 2.7–3.5 eV per Yb atom below the DFT.
• The offset is not a constant. It is a median of 2.75 eV/Yb in intermetallics and 3.3–3.5 in oxides and fluorides, with a 10–90% range of 1.3–5.3. Yb_3 freezes a 4f shell in the core. Do not mix these energies with Yb_3 data or OMat24-trained models through a per-atom shift.

Caveats

• Changed structures. 5,209 replacements changed the structure.

  • 4,046 are the same crystal in a different cell setting (StructureMatcher, ltol 0.02, stol 0.05).
  • 1,163 were re-relaxed by MP (median |ΔV| 2%).
  • 152 of them changed nsites; nsites and chemical_formula_descriptive are updated for those.

  The MLIP energies in LeMat-Bulk-MLIP-Hull-All were computed on the old structures. Recompute them for rows with change == "replace_structure".

• entalpic_fingerprint is unchanged for replaced rows.

• The dft split of LeMaterial/LeMat-Bulk-MLIP-Hull is not corrected here.

• Rows far above all MLIPs are left alone where MP agrees with upstream; they do not distort hulls.

• Upstream has one row with a null immutable_id (Rb, 2 sites). It is kept.

Audit config

audit/changes.parquet:

• change: replace_energy, replace_structure, removed, kept_flagged or forces_stress_added.
• reason: the provenance class.
• cause: the specific evidence, as text.
• Energies: old_energy, new_energy, nsites, new_nsites, old_energy_per_atom and new_energy_per_atom. Total energies are in eV.
• Provenance: old_mp_task, old_calc_type, old_batch, new_mp_task and new_calc_type, in legacy mp- ids.
• Geometry: geometry_match and volume_change.
• forces_source: output or last_ionic_step.
• mlip_median_energy_per_atom: the MLIP median on the upstream structure.

License and citation

CC-BY-4.0, as for upstream. Please cite LeMat-Bulk (doi:10.57967/hf/3762) and the underlying databases: Materials Project for mp- ids, Alexandria for agm ids and OQMD for oqmd- ids. The scripts that produced this dataset are in scripts/.