The Materials Project, OQMD, AFLOW and JARVIS-DFT are the four best-known open databases of density functional theory (DFT) calculations for inorganic crystals. All four compute structures, formation energies and band gaps from first principles, but they differ in who runs them, how many entries they hold, which functional and settings they use, and what license their data carries. Those differences are why the same compound can show slightly different numbers in each one, and why serious screening work usually checks more than one.

• The Materials Project (Lawrence Berkeley National Laboratory, DOE-funded) is widely used, with more than 600,000 users. It mixes GGA, GGA+U and r2SCAN calculations, and its core data is licensed CC BY 4.0.
• OQMD (Wolverton group, Northwestern University) holds about 1.4 million PBE-based entries, with GGA+U for selected elements in oxides. Its data is CC BY 4.0 and the whole database can be downloaded.
• AFLOW (Duke University and the AFLOW consortium) has the most entries, mostly hypothetical prototype decorations. Its data is free for scientific, academic and non-commercial use only.
• JARVIS-DFT (NIST) is smaller but uses a van der Waals functional (OptB88vdW) and adds TBmBJ band gaps, which usually come closer to measured gaps than plain GGA.
• For the same starting structure, databases disagree by amounts comparable to DFT-versus-experiment error. That is a reason to compare sources, not a sign that any one database is wrong.

What is a materials database, and why are these four the reference points?

A computational materials database is a collection of crystal structures run through the same automated DFT workflow, so that properties such as total energy, formation energy, band gap and magnetic moment can be compared across thousands of compounds. Many input structures come from experimental sources such as the Inorganic Crystal Structure Database (ICSD). Hypothetical structures are generated by placing new elements onto known structure types, a process called prototype decoration.

The Materials Project, AFLOW and OQMD grew out of the high-throughput DFT work of the early 2010s, and JARVIS was established at NIST in 2017 (JARVIS overview). The Materials Project's 2013 commentary in APL Materials, the AFLOWLIB.ORG paper (2012), the OQMD paper in JOM (2013) and the JARVIS overview in npj Computational Materials (2020) are the standard citations. All four are registered providers in the OPTIMADE provider registry for the common OPTIMADE query standard, as are newer datasets such as Alexandria. The OPTIMADE specification lets one query reach several of them at once.

How do the Materials Project, OQMD, AFLOW and JARVIS compare at a glance?

Sizes change with every release, and each project counts something slightly different (unique materials, calculation entries or JARVIS IDs), so treat the numbers below as dated snapshots rather than a ranking.

What is the Materials Project?

The Materials Project describes itself as an effort from the U.S. Department of Energy to pre-compute material properties and make them public (MP documentation). It was launched formally in 2011 and is used by more than 600,000 researchers, according to its 2025 Nature Materials perspective. Calculations run in VASP with spin polarization. Elements that need a Hubbard U correction get one in oxides and fluorides, using U values for Co, Cr, Fe, Mn, Mo, Ni, V and W fitted to experimental oxide oxidation enthalpies (a ternary oxide in the case of Ni) (MP Hubbard U documentation). r2SCAN meta-GGA calculations first appeared as pre-release data in database version v2022.10.28 (MP database versions), and MP now combines them with GGA/GGA+U energies using a mixing scheme published by Kingsbury et al. (2022).

Its main strengths are breadth of properties beyond energies (elastic tensors, phonons, electrodes, Pourbaix diagrams), a well-documented API, and a public changelog that records corrections. Note that GNoME-derived structures on MP carry a non-commercial (BY-NC) license that users must accept separately (MP database versions).

What is OQMD?

The Open Quantum Materials Database comes from Chris Wolverton's group at Northwestern. It combines ICSD compounds with a large number of decorated prototypes, which explains its size. Calculations use PBE with VASP PAW potentials, a final static step at a fixed 520 eV cutoff so that energies are comparable, and ferromagnetic initialization for 3d and actinide compounds. The OQMD documentation notes that this approach will not capture antiferromagnetic ground states (OQMD DFT settings).

OQMD's 2015 accuracy study is one of the most useful benchmarks in the field. Comparing 1,670 experimental formation energies, Kirklin et al. found a mean absolute error of 0.096 eV/atom. Different experimental sources disagreed with each other by 0.082 eV/atom, which suggests that a large share of the apparent DFT error comes from the experiments.

What is AFLOW?

AFLOW (Automatic FLOW) is both a software framework and a data repository, developed at Duke University with consortium partners (Curtarolo et al., 2012). Its "AFLOW standard" fixes k-point density, cutoffs, potentials and DFT+U settings so that entries are reproducible. PBE with PAW potentials is the default for the ICSD, binary-alloy and Heusler libraries (Calderon et al., 2015).

AFLOW's very large entry count needs context: most entries are hypothetical. The 2023 aflow.org ecosystem paper lists about 2.6 million entries in the ternary prototype library (LIB3) alone. That breadth is valuable for alloy and phase-stability searches. AFLOW is also unusual in providing automated thermal and elastic workflows (AEL and AGL) and the AFLUX search language. Its license is the main practical constraint: commercial users should read the non-commercial terms carefully.

What is JARVIS-DFT?

JARVIS (Joint Automated Repository for Various Integrated Simulations) is run at NIST and covers DFT, classical force fields and machine-learning models (Choudhary et al., 2020). JARVIS-DFT uses the OptB88vdW functional, which includes van der Waals interactions and so handles layered and two-dimensional materials better than plain PBE. For band gaps it adds the Tran-Blaha modified Becke-Johnson (TBmBJ) potential. The 2018 Scientific Data study found TBmBJ predicts gaps and dielectric functions better than OptB88vdW, at modest cost. JARVIS is the smallest of the four but goes deep on properties such as elastic tensors, topological spillage and optoelectronics.

Why do materials databases disagree on the same compound?

A systematic comparison is Hegde et al. (Physical Review Materials, 2023), who matched AFLOW, MP and OQMD records calculated from the same initial ICSD structure. Differences ran as high as 0.105 eV/atom for formation energy, 0.65 ų/atom for volume, 0.21 eV for band gap and 0.15 μB per formula unit for magnetization. Up to 7% of records disagreed on whether a material is a metal, and up to 15% on whether it is magnetic. The authors traced the larger gaps to choices of pseudopotential, DFT+U treatment and elemental reference energies, and called the spread "comparable to the differences between DFT and experiment."

Energy above hull, the most used stability metric, is especially sensitive to these choices. It is the energy distance from a compound to the lowest-energy combination of competing phases on the convex hull. As the Materials Project glossary explains, it is computed at 0 K. It depends on every other phase in the chemical system and on correction schemes, and any meaningful threshold varies by chemistry. Two databases that contain different competing phases, or correct oxide and transition-metal energies differently, can report different hull distances for an identical structure. Transition-metal oxides such as Fe2O3 are a typical case where the treatment of U matters. Polymorph-rich systems such as TiO2 are another, because small energy differences decide the ordering.

Band gaps add a further layer. GGA-type functionals systematically underestimate gaps, so a PBE gap for Si or GaAs from MP, OQMD or AFLOW will typically sit below both the measured value and JARVIS's TBmBJ value. None of these numbers are measurements. Our explainer on what a band gap is covers why.

Which materials database should you use?

• Battery and electrochemistry screening: the Materials Project, for its calibrated GGA+U oxides, electrode and Pourbaix data, and documented corrections. Cathodes such as LiFePO4 have well-studied entries.
• Phase stability across large composition spaces: OQMD, because the whole database is downloadable under CC BY 4.0 and offers broad hypothetical coverage.
• Alloys, prototypes and thermomechanical properties: AFLOW, provided your use is academic or non-commercial.
• Layered and 2D materials, optoelectronic gaps: JARVIS-DFT, for vdW-aware structures and TBmBJ gaps.
• Anything decision-relevant: check at least two sources and record which functional each value came from.

How does aggregating across databases help?

Aggregation does not average away errors. Its value is that it makes them visible. When independent workflows agree on structure, stability and gap, confidence rises. When they disagree, the size of the disagreement is itself information, often pointing to magnetism, U corrections or missing competing phases. Cross-referencing also catches simple problems such as duplicate entries and mismatched space groups. We discuss this in more depth in what DFT disagreement tells you and in 13 ways computational materials science goes wrong. Aggregation only works if the source projects keep publishing open, well-documented data, and the four discussed here have made that possible.

Frequently asked questions

Is the Materials Project free to use?

Yes. Its core data is licensed CC BY 4.0, which allows reuse with attribution, and it offers a free API. GNoME-derived structures on the site carry a separate non-commercial (BY-NC) license.

Which database is the largest?

By raw entry count AFLOW is largest, at about 3.9 million entries as of September 2026, followed by OQMD at about 1.4 million. Most of these are hypothetical prototype decorations, so entry count is not the same as the number of experimentally known compounds.

Can I use AFLOW data commercially?

AFLOW's published disclaimer states that its data is free for scientific, academic and non-commercial purposes and that any other use is prohibited. Contact the consortium for commercial terms.

Why are JARVIS TBmBJ band gaps often larger than Materials Project gaps?

JARVIS reports TBmBJ gaps for a subset of materials alongside its OptB88vdW gaps, and TBmBJ partly corrects the gap underestimation of GGA functionals such as PBE. Both are still computed values, not measurements.

How accurate are DFT formation energies?

OQMD's benchmark against 1,670 compounds found a mean absolute error of 0.096 eV/atom. Experimental sources disagreed with each other by 0.082 eV/atom, so part of that error lies in the reference data.

Explore the data on LatticeGraph

LatticeGraph cross-references entries from these and other DFT databases on each compound page, labeled by source, with computed stability and band gaps, plus literature synthesis recipes and patents. Start with LiFePO4, TiO2 or GaAs, or browse III-V semiconductors.