---
title: '2DMatPedia: 2D Materials Database'
url: https://www.emergentmind.com/topics/2dmatpedia
type: topic
---

# 2DMatPedia: 2D Materials Database

2DMatPedia is an open computational database of two-dimensional materials designed as a standardized resource for screening, data mining, data analysis, and artificial intelligence applications. It was introduced as a “2D Materials Encyclopedia” containing more than 6,000 monolayer structures with consistently computed structural, electronic, and energetic properties, generated by combining top-down exfoliation of layered bulk compounds with bottom-up chemical substitution. Subsequent work has used the same resource as a database-scale platform for property discovery, including a high-throughput intrinsic spin Hall conductivity evaluation over 4,486 screened non-magnetic compounds drawn from a 6,351-material collection, which effectively adds a transport-property layer to the original structural database [1901.09487][2509.13204].

## 1. Scope, identity, and data model

2DMatPedia is a public database of monolayer materials. The original report presents more than 6,000 monolayer structures, while later database-wide screening work starts from a full 2DMatPedia collection of 6,351 materials, indicating continued growth of the resource after its initial release [1901.09487][2509.13204].

The database is explicitly a property database rather than only a structure repository. For each entry, it stores `material_id`, `relative_id`, `discovery_process` (`top-down` or `bottom-up`), `structure`, `formula`, `nelements`, `elements`, `spacegroup`, `point_group`, `bandgap`, `is_gap_direct`, `is_metal`, `energy_per_atom`, `energy_vdw_per_atom`, `exfoliation_energy_per_atom`, `decomposition_energy_per_atom`, and `total_magnetization` [1901.09487].

This organization makes the resource suitable for both first-principles screening and downstream statistical learning. A direct implication of the stored fields is that 2DMatPedia supports queries spanning symmetry, composition, metallicity, gap character, thermodynamic proxies, and magnetization without requiring a separate preprocessing layer.

| Category | Stored information |
|---|---|
| Identification | `material_id`, `relative_id`, `discovery_process`, `formula` |
| Structure and symmetry | `structure`, `nelements`, `elements`, `spacegroup`, `point_group` |
| Electronic, energetic, magnetic | `bandgap`, `is_gap_direct`, `is_metal`, `energy_per_atom`, `energy_vdw_per_atom`, `exfoliation_energy_per_atom`, `decomposition_energy_per_atom`, `total_magnetization` |

A common misconception is to read the field `discovery_process` as a synthesis history. The database paper states that this terminology refers to how a material is generated in the database, not to an experimental synthesis pathway [1901.09487].

## 2. Database construction: top-down and bottom-up routes

The database is built from two complementary discovery procedures. The top-down route begins from more than 80,000 inorganic compounds in the Materials Project, restricts attention to elemental, binary, ternary, and quaternary compounds with primitive cells containing fewer than 40 atoms, and applies a topology-based algorithm to identify layered structures [1901.09487].

The layeredness test uses the standard conventional unit cell, defines bonds through a covalent-radius threshold, groups bonded atoms into clusters, builds a \(3\times 3\times 3\) supercell, and labels a structure as layered when the number of clusters in the supercell is three times that in the unit cell. This test is repeated over the tolerances
\[
0.0,\ 0.05,\ 0.1,\ 0.15,\ 0.2,\ 0.25,\ 0.3,\ 0.35,\ 0.4
\]
and only structures identified as layered by at least two tolerances are retained. For each layered bulk, one cluster is extracted and a vacuum spacing greater than \(20\ \text{\AA}\) is added along the \(c\)-axis to construct the monolayer model [1901.09487].

The bottom-up route expands the search space beyond layered bulk parents. Starting from known 2D materials, it groups elements by periodic-table column, excludes radioactive elements and lanthanides and actinides with \(f\) electrons except La, and performs chemical substitution within the same group. The database paper gives BN as an illustrative case: replacing B with \([B, Al, Ga, In, Tl]\) and N with \([N, P, As, Sb, Bi]\) produces 24 new materials [1901.09487].

The reported counts are 2,884 materials from the top-down route and 2,927 materials from the bottom-up route. This suggests that the database was designed not only to recover exfoliable monolayers but also to enumerate chemically analogous hypothetical compounds that may be metastable yet still synthesizable [1901.09487].

## 3. Computational methodology and descriptors

The original database calculations follow the Materials Project standard high-throughput DFT workflow using VASP, PBE, the PAW method, a plane-wave cutoff of 520 eV, and the vdW-optB88 functional. For 2D materials, both cell shape and volume are relaxed while the \(c\)-axis is fixed, with ionic convergence threshold
\[
\Delta E = 1.0 \times 10^{-5}\ \text{eV}.
\]
A static calculation with a uniform \(\Gamma\)-centered \(k\)-point grid generates the charge density, after which non-self-consistent DOS calculations are performed over \(E_F - 10\ \text{eV}\) to \(E_F + 10\ \text{eV}\) with 2000 intervals; band structures use a line-mode path through high-symmetry points of the 2D Brillouin zone, and vdW correction is not applied in DOS or band-structure runs [1901.09487].

The energetic descriptors are central to how 2DMatPedia is used. Exfoliation energy is defined as the average energy per atom required to remove a layer from its layered bulk parent, with the lowest-energy parent selected when multiple bulk counterparts exist. For bottom-up materials with no real bulk parent, a constructed layered bulk is formed by stacking the monolayers in the original sequence. Decomposition energy is defined as a modified form of energy above hull that excludes the layered bulk parent from competing phases and is obtained from a phase-diagram construction using 2D material energies without vdW correction for direct comparison [1901.09487].

The paper explicitly notes that, for special cases such as silicene, exfoliation energy is only a reference metric and decomposition energy better reflects stability. This is important because it clarifies that kinetic accessibility and substrate stabilization are not reducible to a single energetic scalar, and it explains why the database intentionally retains less-stable entries [1901.09487].

## 4. Validation, public access, and intended use

Technical validation is reported against both an external database and experimentally realized materials. Relative to the JARVIS 2D database, the mean error in \(a\) is \(0.0090\ \text{\AA}\) with standard deviation \(0.034\), the mean error in \(b\) is \(0.010\ \text{\AA}\) with standard deviation \(0.062\), and the exfoliation-energy difference has mean \(-1.13\ \text{meV/atom}\) with standard deviation \(5.64\ \text{meV/atom}\); 93% of the 383 compared data points have exfoliation-energy differences smaller than \(10\ \text{meV/atom}\) [1901.09487].

Against experimentally grown 2D materials, 53 of 59 analyzed cases have decomposition energy below \(100\ \text{meV/atom}\), and 39 are within \(10\ \text{meV/atom}\). The database paper interprets this as evidence that decomposition energy is a useful indicator of synthetic accessibility and stability, while also noting known metastable-but-real materials such as silicene, germanene, monolayer MoO\(_3\)-related phases, and T’-MoS\(_2\) [1901.09487].

The database is publicly available at `http://www.2dmatpedia.org/`, with a downloadable JSON file at `http://www.2dmatpedia.org/download`. The implementation is built using Materials Project tools, especially `pymatgen 4.7.3` and `atomate 0.5.1` [1901.09487].

Its intended applications are stated explicitly: materials screening, data mining, data analysis, and artificial intelligence or machine learning. A plausible implication is that the uniformity of the DFT workflow is as important as the database size, because cross-material comparison and supervised learning are both sensitive to workflow-induced label heterogeneity.

## 5. Extension into a high-throughput spin-transport platform

A major later development is the database-wide intrinsic spin Hall conductivity screening reported in “High-throughput screening of spin Hall conductivity in 2D materials” [2509.13204]. That work begins from the full 2DMatPedia collection of 6,351 materials, removes 1,865 magnetic 2D materials, and evaluates spin Hall conductivity for 4,486 non-magnetic compounds. Its main contribution to 2DMatPedia is not new structures, but a transport-property layer that ranks existing entries for spintronics [2509.13204].

The workflow is automated and scale-oriented. Starting structures are taken from 2DMatPedia; electronic structure is computed with VASP using the PAW method, PBE-GGA, and SOC, with an I-centred \(k\)-mesh density of \(0.02\ \text{\AA}^{-1}\) and a 550 eV plane-wave cutoff. An in-house automated wannierization workflow integrates VASP with Wannier90 to construct MLWF-based tight-binding Hamiltonians, and SHC is evaluated with WannierBerri on a dense \(400 \times 400 \times 1\) Brillouin-zone mesh with adaptive refinement [2509.13204].

The screening reveals a broad SHC distribution in which most compounds are below about \(200\,(\hbar/e)(\mathrm{S/cm})\), but 27 materials have SHC \(>300\,(\hbar/e)(\mathrm{S/cm})\) and 6 exceed \(500\,(\hbar/e)(\mathrm{S/cm})\). The highest-profile entries are Bi\(_4\)Pb\(_3\), Te\(_4\)Os\(_3\), Ge\(_3\)Bi\(_4\), Sn\(_3\)Bi\(_4\), Tl\(_3\)Os\(_2\), and PtPb\(_3\), with Bi\(_4\)Pb\(_3\) reported as the highest-SHC material. The same study also identifies 57 two-dimensional topological insulators with quantized SHC [2509.13204].

The physical interpretation extracted from this screening is that large intrinsic SHC is associated with strong SOC in heavy-element systems, SOC-induced gap opening at Dirac-like crossings near the Fermi level, and mirror symmetry. Among materials with SHC greater than \(100\,(\hbar/e)(\mathrm{S/cm})\), 71.6% have at least three mirror planes; this rises to 76.8% for materials above \(200\,(\hbar/e)(\mathrm{S/cm})\). The study therefore converts 2DMatPedia from a catalog of 2D structures into a screened spin-transport resource [2509.13204].

## 6. Role in adsorption ML and closed-loop inverse design

2DMatPedia has also been used as a low-data benchmark for adsorption-energy prediction and as a seed pool for closed-loop inverse design. In “Local environment-based machine learning for molecular adsorption energy prediction,” the authors describe 2DMatPedia as containing more than 10,000 distinct 2D materials and use a subset of 2,472 DFT hydrogen adsorption calculations. Their Local Environment ResNet is transferred from the larger OC20 benchmark to this 2DMatPedia subset, where it is reported to maintain stable performance superiority under low-data conditions. The study further states that 1,283 adsorption sites were generated and predicted from 2DMatPedia, yielding 272 predicted adsorption energies within \(\pm 0.5\) eV, including 7 reported experimentally, 69 from other computational studies, and 196 previously unreported in the public literature or open databases [2311.11364].

In “InvDesMobility: a reliability-gated first-principles feedback framework for closed-loop materials discovery,” 2DMatPedia-derived 2D semiconductor candidates are the seed pool for a carrier-mobility campaign. That workflow starts from 516 candidates selected after structural and electronic pre-screening, attempts 2,064 in-plane mobility channels, reports 280 QC-passed materials, 83 caution-level materials, and 153 not retained, and preserves 573 retained seed feedback channels for learning. Over multiple iterations, the loop screens \(2.4\times 10^6\) structures, submits 102 candidates for DFT validation, and retains 86 reliability-gated generated channels across 41 formulas [2606.16133].

These downstream uses clarify the position of 2DMatPedia within the computational materials ecosystem. It functions simultaneously as a monolayer-structure repository, a source of consistently computed descriptors, a seed set for physics-based high-throughput screening, and a substrate for data-efficient ML and auditable inverse-design workflows. This suggests that the database’s long-term significance lies not only in its original enumeration of 2D materials, but also in its reusability as a standardized prior for successive layers of property inference and experimental prioritization [2311.11364][2606.16133].

Source: https://www.emergentmind.com/topics/2dmatpedia