benchmarks/materials/c2db_2d_property/data/README.md

C2DB 2D-property data file

Usage

bash download.sh          # runs the scraper; writes data.csv and schema.json

The scrape walks the live C2DB browse UI at https://c2db.fysik.dtu.dk/ (the 2DHub redirect target) page by page. The run takes ~8–12 minutes end to end; progress prints to stdout. Re-running the script is idempotent (it overwrites data.csv).

data.csv is gitignored per the repo-wide rule (benchmarks/**/data/data.* in .gitignore) — you always rebuild it locally from this script. Only this README.md and the optional schema.json are kept in git.

Source and license

• Upstream landing: https://2dhub.org/c2db/c2db.html
• Browse UI (what we scrape): https://c2db.fysik.dtu.dk/
• DTU Figshare record: https://doi.org/10.11583/DTU.14616660.v1
• License: CC-BY-4.0 (declared on the Figshare record).
• Primary citation: Haastrup et al., 2D Mater. 5, 042002 (2018), https://doi.org/10.1088/2053-1583/aacfc1; update Gjerding et al., 2D Mater. 8, 044002 (2021), https://doi.org/10.1088/2053-1583/ac1059.

When redistributing or publishing work that uses this data.csv, cite the Haastrup and Gjerding papers (both are vendored offline in ../reference/).

Filter applied

The C2DB browse UI ships a server-side default filter ehull in [0.0, 0.2] eV/atom and rejects wider ranges with HTTP 500 (tried: to_ehull = 0.5, 1.0, 2.0, 5.0, 10.0 — all 500). The scraper therefore collects the 9706 visible rows out of the 17001 total C2DB entries. This is the same "energetically accessible" subset the official web UI shows to anyone clicking "Browse data"; the remaining ~7000 entries are above 0.2 eV/atom above the convex hull and likely not synthesizable.

If you need the full 17k entries, you have to contact the C2DB authors ("The full dataset is provided upon request", per 2dhub.org). This entry deliberately does not try to back-door that process.

Column order in data.csv

Column 0 is the primary output (hform, heat of formation eV/atom). Columns 1..N are inputs, in this order:

1. hform — output — Heat of formation [eV/atom] (PBE)
2. uid — C2DB unique material identifier (string)
3. formula — chemical formula (string)
4. layergroup — 2D layer group symbol (string)
5. is_magnetic — Magnetic Yes/No
6. dyn_stab — Dynamically stable Yes/No/Unknown

Then every other scalar column the C2DB web UI exposes, in alphabetical order. Depending on the live schema at build time, this includes:

key	description	unit
A	Single-ion anisotropy (out-of-plane)	meV
E_B	Exciton binding energy (BSE)	eV
J	Nearest-neighbour exchange coupling	meV
N_nn	Number of nearest neighbours	(int)
Px / Py / Pz	Spontaneous polarization components	pC/m
P_spontaneous_norm	Spontaneous polarization norm	pC/m
alphax / alphay / alphaz	Static polarizability (ph + el)	Å
alphax_el / alphay_el / alphaz_el	Interband polarizability	Å
alphax_lat / alphay_lat / alphaz_lat	Phonon polarizability	Å
area	2D unit-cell area	Ų
bravais_type	2D Bravais type (e.g. Hexagonal (hp))	(text)
cbm / cbm_evac / cbm_u / cbm_hse / cbm_gw	CBM vs vacuum (PBE/PBE+U/HSE/GW)	eV
cod_id	Parent bulk structure in COD	(int)
dE_zx / dE_zy	Magnetic anisotropy energy	meV/unit cell
dipz	Out-of-plane dipole moment	e·Å/unit cell
efermi / efermi_u / efermi_hse / efermi_gw	Fermi level vs vacuum	eV
ehull	Energy above convex hull	eV/atom
emass_cbm / emass_vbm	DOS effective masses (PBE)	m_e
energy	PBE total energy	eV
evac / evacdiff	Vacuum level and vacuum-level shift	eV
gap / gap_u / gap_hse / gap_gw	Indirect band gaps	eV
gap_dir / gap_dir_u / gap_dir_hse / gap_dir_gw	Direct gaps	eV
halfmetal_gap / halfmetal_gap_hse / halfmetal_gap_dir / halfmetal_gap_dir_hse	Minority-spin gaps	eV
has_inversion_symmetry	Inversion symmetry flag	(bool)
icsd_id	Parent bulk structure in ICSD	(int)
is_ferroelectric	Ferroelectric flag	(bool)
is_magnetic_u	Magnetic flag (PBE+U)	(bool)
label	Origin label (e.g. Original18, ICSD-derived)	(text)
lam	Anisotropic exchange (out-of-plane)	meV
lgnum	Layer group number	(int)
magmom	Total magnetic moment	μ_B
natoms	Number of atoms per unit cell	(int)
nspecies	Number of distinct chemical species	(int)
number	Bulk space group number (AA-stacked)	(int)
plasmafrequency_x / plasmafrequency_y	Drude plasma frequency	eV·Å^0.5
spin	Max S_z at magnetic sites	(dim-less)
thickness	2D layer thickness (out-of-plane extent)	Å
vbm / vbm_evac / vbm_u / vbm_hse / vbm_gw	VBM vs vacuum (PBE/PBE+U/HSE/GW)	eV
workfunction	Work function (averaged for polar slabs)	eV

Text-only / internal columns are intentionally skipped by the download script: doi (free-text citation), publication (free-text), folder (internal DTU path), olduid (legacy id), stoichiometry (redundant text), reduced (redundant text), international / pointgroup (text symbols). The exact runtime column set is saved to schema.json by download.sh for auditing.

Missing values

A cell is empty ("") whenever that property is undefined for a given material or simply wasn't computed in the run the web UI serves:

• gap_gw / gap_dir_gw / vbm_gw / cbm_gw / efermi_gw / E_B — computed only for the ~300-material GW/BSE subset.
• workfunction / efermi — often blank for metals, where vacuum-alignment is ambiguous in the presence of dipoles.
• emass_cbm / emass_vbm — undefined for metals / half-metals.
• J / lam / A / dE_zx / dE_zy / magmom — blank for non-magnetic materials.
• Px / Py / Pz / P_spontaneous_norm — blank for non-ferroelectric materials.
• cod_id / icsd_id — blank for materials derived from generative models rather than from ICSD/COD parents.
• Many properties (especially optical / phonon) may be blank for newer, partially-computed entries (the 2022-09 generative-model additions and the 2024-05 restructuring).

Pandas users: pd.read_csv('data.csv', na_values=[''], keep_default_na=True) is the recommended reader.

Known build-time quirks

• The sid session id in the scraper URL is server-side state; it gets reallocated on every GET / hit. The script creates exactly one sid and sticks with it.
• Toggle requests are sent with the HX-Request: true header to match the htmx client the UI was built around. Server-side toggling is purely per-session, so concurrent scrapers would interfere with each other on shared sids — do not parallelize this script.
• A 0.35 s sleep between page GETs keeps the rate well below what the CAMD web server serves to an interactive user; this is intentional.