The Same Perovskite, Different Stability Verdicts

Query SrTiO3 in Materials Project and OQMD, and you will get different hull distances. Not because one database made an error. Both did the calculation correctly. The differences come from parameter choices that are individually defensible but produce inconsistent numbers when compared directly. A 2023 study in Physical Review Materials (Kingsbury et al., PRM 7, 053805) found that up to 7% of compounds across AFLOW, MP, and OQMD disagree on whether a material is metallic, and up to 15% disagree on whether it is magnetic. For perovskites with transition-metal B-sites (the majority of technologically relevant oxide perovskites), these rates are the floor, not the ceiling.

This article explains the specific sources of disagreement and gives a concrete protocol for handling them in a screening workflow.

What Perovskite Stability Metrics Actually Measure

Formation Energy

Formation energy is E(compound) minus the sum of elemental reference energies. That second term is where databases diverge. Each database computes its elemental reference states independently, using its own DFT settings. Even with the same functional and the same compound energy, different elemental references shift all formation energies by a systematic offset. You cannot subtract a Materials Project formation energy from an OQMD formation energy and interpret the difference as physically meaningful.

Energy Above Hull

Energy above hull measures thermodynamic stability relative to all competing phases in the database. It depends directly on which competing phases are included in the convex hull construction. A database with more computed phases defines a different hull shape. The same compound can show a different hull distance in each database not because the compound energy differs, but because the competing phase space differs.

Goldschmidt Tolerance Factor as a Pre-Filter

The Goldschmidt tolerance factor (t between roughly 0.75 and 1.0 for stable perovskite structures) is a geometry-based pre-filter independent of DFT functional and database choice. Use it to screen candidate compositions before running any DFT-derived stability check.

Formation energy is functional-dependent and reference-dependent. Hull distance is additionally phase-space-dependent. Neither is an absolute property of the material.

Why the Four Databases Disagree on Perovskite Energetics

Functional Choice and DFT+U Parameters

Materials Project and OQMD apply Hubbard U corrections systematically for transition metal oxides, using independently-calibrated U values for specific elements. AFLOW’s database is primarily GGA-PBE, with U corrections applied selectively rather than as a blanket policy. For perovskite B-site ions (Ti, Fe, Mn, Co, Ni), the U value directly affects the total energy. MP’s U(Fe) is 5.3 eV for oxides. Independently-calibrated values across databases can differ by approximately 1 eV, which shifts iron-containing formation energies enough to change relative stability rankings at the margins.

JARVIS-DFT uses the OptB88vdW functional throughout, without the GGA+U scheme. This is not a small adjustment. It is a different energy reference frame. OptB88vdW was chosen because it gives accurate lattice parameters for both van der Waals and non-vdW solids. For oxide perovskites (which are not van der Waals bonded), this choice introduces a systematic energy offset relative to the PBE+U results in the other three databases.

Elemental Reference Energies

Each database computes its own elemental reference states, and the choices matter: magnetic ground state for elemental iron, structural phase for elemental tin, DFT settings for molecular oxygen. Small differences combine when formation energy is calculated as E(compound) minus elemental sum. If two databases agree on the compound energy but use different oxygen reference energies, their oxide formation energies will differ by a systematic offset proportional to the oxygen stoichiometry. That shift applies to every oxide in the database.

Convex Hull Phase Space Completeness

Materials Project has the most comprehensive phase space coverage for most oxide chemical systems, which generally makes its hull distances the most reliable for identifying metastable compounds. OQMD covers oxide chemistry well and is a strong second source for cross-validation. AFLOW has the most entries (~3.93M) but hull completeness for specific ternary systems varies. JARVIS constructs its hull from ~40,000 structures, which limits competing phase coverage for complex perovskite compositions.

Three independent sources of disagreement (functional and U values, elemental references, and phase space coverage) combine for perovskites with transition-metal B-sites. The disagreements are systematic and predictable, not random.

JARVIS-DFT: Accurate Structures, Different Energy Scale

OptB88vdW vs. GGA-PBE+U for Non-vdW Perovskites

JARVIS chose OptB88vdW because it produces better structural parameters than GGA-PBE for a wider range of materials, including systems where dispersion matters. For perovskite lattice constant prediction, JARVIS frequently produces closer agreement with experiment than the PBE+U databases. That structural accuracy is worth using in workflows where lattice parameters feed into property predictions or structure-property models.

But do not mix JARVIS formation energies directly with MP or OQMD values. The functional difference creates a systematic energy offset that is not a constant correction factor. It varies by chemistry. Relative stability rankings within JARVIS are internally consistent and valid. Absolute formation energy comparisons across the OptB88vdW/PBE+U boundary require functional-aware corrections that have not been standardized for the full perovskite compositional space.

K-Point Convergence Strategy

JARVIS uses per-material k-point convergence rather than a fixed k-point density scheme. Starting from the Γ point and increasing k-point density until energy convergence is confirmed, this approach can produce different total energies for materials where other databases’ k-grids are insufficiently dense. For most cubic perovskites, this is unlikely to be a significant factor. For low-symmetry distorted perovskites, it is worth checking.

JARVIS sources its input structures from Materials Project, AFLOW, OQMD, and ICSD. Use the JVASP identifier to search for corresponding entries in the other databases before comparing stability metrics. JARVIS values are internally consistent and useful for structural accuracy and electronic properties. They sit on a different energy scale than the PBE+U databases.

Why Magnetic Perovskites Are the Hardest Case

Spin Configuration and U Value Sensitivity

LaFeO3, LaMnO3, SrMnO3, and similar transition-metal oxide perovskites are where cross-database disagreements peak. Two factors combine: different U values for the magnetic B-site ion, and different spin configurations used as the DFT starting point. If one database initializes with ferromagnetic ordering and another with antiferromagnetic ordering, they can converge to different local energy minima. Both results are self-consistent within their respective calculations, but they represent different magnetic ground states with different total energies.

Kingsbury et al. 2023 found up to 15% of compounds across AFLOW, MP, and OQMD disagree on whether a material is magnetic. For a perovskite screening workflow that includes iron or manganese B-sites, this means a meaningful fraction of candidates will carry different stability verdicts in different databases.

Compounds Most Affected

The highest disagreement risk concentrates in perovskites where the B-site carries a significant magnetic moment: Fe3+ (LaFeO3, BiFeO3), Mn3+ and Mn4+ (LaMnO3, SrMnO3, mixed manganites), Co3+ (LaCoO3), and Ni3+ (LaNiO3). Non-magnetic perovskites (SrTiO3, BaTiO3, SrZrO3) show considerably smaller cross-database formation energy differences.

Magnetic perovskites should never be screened against a single database. Cross-validation against at least one additional source is the defensible minimum for any screening study feeding into synthesis decisions.

Which Database for Which Perovskite Use Case?

Database-by-Database Guidance

Start with Materials Project for hull distance assessment. It has the most comprehensive competing phase coverage for most oxide systems, and its hull distances are the most reliable starting point for thermodynamic stability screening. Check whether your target compositions have r2SCAN data available in MP; the r2SCAN functional migration adds more accurate energetics for a growing entry subset.

Use OQMD as the primary cross-validation source for oxide perovskites. Full GGA-PBE consistency across the entire database makes it the best second opinion on formation energy. For any compound where MP and OQMD hull distances differ by more than roughly 0.05 eV/atom (a practical threshold based on typical DFT functional errors for oxides), treat the stability verdict as uncertain and investigate the specific competing phases each database includes.

Use AFLOW for breadth. Ternary and quaternary perovskite systems where MP coverage thins out are where AFLOW’s ~3.93M entries earn their value. Formation energy trends and descriptor-based screening across large composition spaces are AFLOW’s strongest use case.

Use JARVIS for lattice parameter accuracy, optical and electronic properties, and any system where dispersion interactions are relevant (2D perovskites, Ruddlesden-Popper phases). Use the JVASP identifier to locate the same compound in MP or OQMD before comparing stability metrics. Do not mix raw energies across the OptB88vdW/PBE+U boundary.

A Practical Cross-Database Validation Protocol

A defensible screening protocol for oxide perovskites:

  • Pre-filter by Goldschmidt tolerance factor (t between 0.75 and 1.0).

  • Screen formation energy in Materials Project, using r2SCAN where available.

  • Cross-check hull distance in OQMD for candidates that pass.

  • Flag any compound where MP and OQMD disagree on the sign of the hull distance. Those require individual review of which competing phases each database includes.

  • For magnetic B-site compositions (Fe, Mn, Co), add a JARVIS bandgap check as a third data point on electronic structure.

When running these queries, Alloybase hits all four OPTIMADE-compliant providers in a single request. You choose which compounds to run the cross-provider comparison on; disagreements appear in your dataset rather than requiring manual cross-referencing across four separate interfaces. Start free at alloybase.app, no credit card required.

No single database is authoritative for perovskite stability. A two-database minimum with functional-aware comparison is the defensible standard for any screening study that will feed into a publication or synthesis campaign.

FAQ

Can I average formation energies across databases?

No. Formation energies from different functionals and reference states are not on the same energy scale. Averaging them combines systematic biases rather than reducing random error. Compare within a single database, or apply explicit corrections before any cross-database arithmetic. The Kingsbury et al. 2023 study (PRM 7, 053805) is the most rigorous published analysis of what those systematic differences look like quantitatively.

Which database should I trust for hull distance?

Materials Project for most oxide perovskite systems, because it has the most comprehensive competing phase coverage. But verify that the specific chemical system you are screening is well-represented in MP. For systems with sparse MP coverage, OQMD is the next best option for hull distance reliability.

How do I compare JARVIS values with MP or OQMD?

Use the JVASP identifier to locate the same compound in MP or OQMD. Do not subtract JARVIS formation energies from MP or OQMD values directly. The OptB88vdW and PBE+U functionals use different energy reference frames, and the offset is not a constant correction factor. Relative stability rankings within JARVIS are valid. Absolute cross-database arithmetic requires functional-aware corrections.

Do these disagreements matter for high-throughput screening?

For clearly stable compounds (hull distance well below 0.1 eV/atom), databases usually agree and a single-source screen is adequate. For candidates near the hull boundary (0 to 0.05 eV/atom), the disagreements are large enough to flip a pass/fail decision. Any metastable candidate feeding into a synthesis decision should be validated in a second database.