Our discovery pipeline scans public computational databases for unusual materials. This week it kept pointing at a few suspicious records. Twice, the problem was not in the material but in the database itself. Here is what we found, how we checked it, and where we reported it.
The two databases
Both are run by U.S. national laboratories and are among the most widely used reference datasets in computational materials science.
| Database | Operated by | What it is |
|---|---|---|
| Materials Project | Lawrence Berkeley National Laboratory, U.S. Department of Energy | Computed properties of more than 150,000 inorganic materials, used by 600,000+ registered researchers. Its energy corrections live in pymatgen, the open-source library maintained with the project. |
| JARVIS-DFT | National Institute of Standards and Technology (NIST), U.S. Department of Commerce | DFT-computed structures and properties, including piezoelectric, dielectric and infrared responses, for tens of thousands of materials. |
What we were doing
We are building an autonomous materials-discovery harness. It scans public databases for materials whose computed properties or stability look out of line, and then tries to prove each lead wrong before believing it. A surprising number counts as a finding only after three checks:
- It matches the raw calculation files and the database's own paper.
- A negative control shows nothing changes where nothing should.
- An independent calculation reproduces the key number.
Both issues below came out of that routine.
1. JARVIS-DFT: the piezoelectric field holds only half the physics
Our piezoelectric screen ranked materials by the downloadable dfpt_piezo_max_eij field of NIST's JARVIS dft_3d dataset. Well-known piezoelectrics came out far below the values the JARVIS team reports on its own website and in its paper.
| Material | JARVIS ID | Paper, Table S2 (C/m²) | Dataset field (C/m²) |
|---|---|---|---|
| BaTiO₃ | JVASP-110 | 4.13 | 0.371 |
| ZnO | JVASP-1195 | 1.10 | 0.787 |
| AlN | JVASP-39 | 1.39 | 0.419 |
| LiNbO₃ | JVASP-1240 | 1.59 | 0.623 |
| GaN | JVASP-30 | 0.47 | 1.043 |
| SiO₂ | JVASP-41 | 0.16 | 0.159 |
How we verified it. We opened the raw DFPT output (VASP OUTCAR) that JARVIS links for ZnO.
- The electronic (clamped-ion) block gives a maximum |e| of 0.78746, exactly the stored value.
- Adding the ionic contribution gives 1.104, which matches the JARVIS web page and the paper.
So the downloadable field most likely holds only the electronic part of the tensor. A likely cause is a parser that labels the electronic block as the total. One case (GaN) does not fit this explanation, and we said so.
Reported: GitHub issue usnistgov/jarvis #340 (27 September 2026), with the comparison table, the raw-file check and the suspected code location.
2. Materials Project: anion energy corrections silently dropped
While screening oxyhalides, compounds chemists have made for decades appeared strongly unstable in Berkeley Lab's Materials Project. NbOCl₃, for example, sat 0.343 eV/atom above the convex hull, far outside the range of anything that should exist.
The Materials Project applies empirical energy corrections to anions (O, N, Cl, Se and others). Its code in pymatgen looks up each anion's oxidation state with a string key such as "Cl". The stored corrections match exactly what happens when the correct oxidation states are passed with Element-object keys instead. The lookup returns 0, and only the most electronegative anion is corrected.
oxidation_state = entry.data["oxidation_states"].get(anion, 0) # "Cl" not found -> 0For NbOCl₃, recomputing with string keys restores the missing chlorine correction (−7.37 eV per cell). With Element keys the result is identical to what the database stores.
| Measure | Result |
|---|---|
| Entries whose stored corrections we reproduced exactly | 152,426 / 152,426 |
| Entries missing at least one anion correction | 10,885 (N 3,487 · Cl 2,172 · Se 1,394 · Te 1,307 …) |
| Materials that move onto the hull once corrected | 1,560, of which 1,214 are experimentally known |
Examples, stored → corrected energy above hull (eV/atom): NbOCl₃ 0.343 → 0 · TaOCl₃ 0.355 → 0 · FeOCl 0.178 → 0 · BiOCl 0.040 → 0 · PbFCl 0.182 → 0 · TaON 0.087 → 0 · LaTaON₂ 0.140 → 0 · Li₃OCl 0.150 → 0.028.
How we verified it.
- Mechanism: the
Element-key recomputation matches every stored correction in the database. - Negative control: 33,057 of 33,057 anion-free entries are unchanged.
- Independent DFT: we ran our own Quantum ESPRESSO calculation (PBE, all phases relaxed) of 5 NbOCl₃ → Nb₂O₅ + 3 NbCl₅. The anion counts balance on both sides, so the corrections cancel and plain DFT can be compared directly. NbOCl₃ comes out at least 0.026 eV/atom below the decomposition products, i.e. stable, as experiment says. The Materials Project's own r2SCAN value (0.006 eV/atom) points the same way.
This bug was first noticed by another user on the Materials Project forum in August 2026, for an oxynitride. Our contribution is the database-wide scope, the finding that it is not limited to mixed-anion compounds, and the independent verification.
Reported: GitHub issue materialsproject/pymatgen #4706 (28 September 2026), with a minimal reproduction, the full scope, a suggested one-line fix, and credit to the original forum report.
Why this matters
Databases run by Berkeley Lab and NIST are shared infrastructure for the whole field. Researchers use them to pick candidates for synthesis, to benchmark methods, and to train the machine-learning models that now drive much of computational materials discovery.
| Indicator | Figure |
|---|---|
| Registered Materials Project users | 600,000+ (Berkeley Lab, 2025) |
| Citations of the Materials Project paper (Jain et al., 2013) | 13,297 |
| Citations of the pymatgen paper (Ong et al., 2013) | 4,530 |
| Citations of the JARVIS paper and the JARVIS DFPT/piezoelectric paper | 538 and 127 |
We cannot count how many studies actually used the affected records, and we do not claim that any particular result is wrong. But the potential reach is large.
- Work that used Materials Project stability for oxyhalides, oxynitrides, BiOX photocatalysts, PbFCl-type compounds or antiperovskite electrolytes may have seen known materials labelled unstable.
- A model trained on the JARVIS piezoelectric field learned only the electronic part.
Fixing these at the source helps everyone downstream at once.
What we take from it
Anomaly detection over public data finds data errors as readily as it finds new materials. So we now treat every downloaded field as a claim to be checked against raw files, the source paper and an independent calculation before it enters our pipeline. When the check fails, we report upstream with a minimal reproduction and credit anyone who found it first.
Looking for researchers and labs ready to put AI to work
Let AI handle candidate search and computational checks, so your team can focus on the questions that matter.
- Synthesis & characterization labs: design synthesis runs around AI-screened, pre-verified candidates, with fewer dead ends and faster discoveries.
- Bio & life-science researchers: genomes, proteins, therapeutic targets. Tell us the problem you need solved, and we'll find the answer together.
- Database builders & users: if you build or rely on scientific data, we'll help you audit it for hidden errors, the same way we did here.
Write to us at discovery@ai3.kr. Two lines about your work are enough to start.
Citation counts from OpenAlex, retrieved 28 September 2026. Materials Project user figure from Berkeley Lab (2025). Both issues were open and awaiting maintainer response at the time of writing.