Fetching external data with vqfetch

You’ll learn: how to pull crystal structures from open databases (OPTIMADE / Materials Project / COD) and experimental reference data from NIST CCCBDB into ready-to-run vibe-qc inputs, with full per-record provenance (source ID + URL + DOI + license + fetched timestamp) baked into every emitted file.

Why: stop hand-transcribing lattice constants from papers and atomization energies from CCCBDB into your input files. The fetcher does it for you, attaches the citation, caches the result, and emits a regression-suite SPEC + an executable input script. Provenance is in the artefact, reproducibility by construction, not retroactively.

Prerequisites: vibe-qc installed with the fetch extra:

pip install -e '.[fetch]'   # adds optimade-python-tools, ase, bs4, lxml

Two runnable example scripts ship alongside this tutorial:

Part 1: pull a crystal structure (Phase 1)

The shortest path to a ready-to-run vibe-qc input is the canonical set, five round-trip-verified structures with pinned Materials Project IDs:

vqfetch canonical mgo_rocksalt --quick

This prints exactly two paths to stdout (both emitted at runtime, not checked into the repo):

examples/regression/systems/periodic/mgo_rocksalt.py
examples/input-mgo_rocksalt-sto-3g.py

The first is a regression-suite PeriodicSpec module; the second is a standalone executable that calls vq.run_rks_periodic_scf on the fetched cell. Run the second one:

.venv/bin/python examples/input-mgo_rocksalt-sto-3g.py

Every file emitted carries Provenance:

provenance=Provenance(
    source_db='OPTIMADE/mp',
    source_id='mp-1265',
    source_url='https://optimade.materialsproject.org/structures/mp-1265',
    original_reference='',
    license='CC-BY-4.0',
    fetched_at='2026-05-16T16:16:34Z',
    fetcher_version='0.1.0',
    notes='',
)

Programmatic equivalent

The same fetch-and-emit flow is available from Python. This looks up the rocksalt MgO structure by its Materials Project ID, then writes both the regression-suite spec module and a runnable input script:

from vibeqc.fetch import (
    emit_input_script, emit_spec_module, fetch_optimade,
)

# ID-lookup is unambiguous — formula-based queries hit the
# polymorph-disambiguation footgun (MgO has rocksalt + CsCl + hex
# polymorphs all in MP; see § 11 of the structure-fetcher handover).
spec = fetch_optimade(optimade_id="mp/mp-1265", quick=True)[0]

emit_spec_module(spec,  "examples/regression/systems/periodic/")
emit_input_script(spec, "examples/", basis="sto-3g", method="rks-lda")

For when you only have a formula and want to see what polymorphs exist, the multi-candidate API (v0.8.0) returns a deduped ranked list:

candidates = fetch_optimade(formula="MgO", max_results=5)
for s in candidates:
    print(f"  sg={s.space_group}  a={s.lattice_ang[0][0]:.3f}  "
          f"n_atoms={len(s.atoms)}  {s.provenance.source_id}")

Output (across the OPTIMADE federation):

  sg=Fm-3m  a=4.194  n_atoms=8  mp-1265        ← canonical rocksalt
  sg=Pm-3m  a=2.661  n_atoms=2  mp-1009127     ← HP CsCl phase
  sg=P6_3mc a=3.27   n_atoms=4  mp-1191789     ← HP hexagonal phase
  …

docs/roadmap.md § VFETCH-X1 tracks the vqfetch list-candidates CLI surface for the same workflow (v0.8.x maintenance line).

Part 2: pull experimental reference data (Phase 2)

The vqfetch reference command pulls experimental data for a molecule by its CAS number from NIST CCCBDB. Here it fetches the water record:

vqfetch reference --cas 7732-18-5      # H₂O — 7732-18-5 is its CAS

Emits a JSON ExperimentalReference record under examples/regression/references/<cas>.json containing atomization energy, ΔH_f at 0K and 298K, vibrational fundamentals

  • harmonics, IR intensities, ionization energy, proton affinity, dipole moment, polarizability, bond list, and the explicit Cartesian geometry, all sourced from CCCBDB’s exp2x.asp page, all carrying the NIST DOI doi:10.18434/T47C7Z in Provenance.

⚠ The D₀ vs Dₑ footgun

CCCBDB’s atomization energy is D₀, the thermodynamic value at 0 K, with zero-point vibrational energy already included. The value most QC textbooks quote (and the value most QC method papers report) is Dₑ, the electronic-only atomization energy, ZPE removed. These differ by +ZPE = +½·Σν, which is not small (≈ 13 kcal/mol for water, ≈ 28 kcal/mol for CH₄).

For water:

AE D₀ = 219.35 kcal/mol  (CCCBDB)
+ZPE  =  12.88 kcal/mol  (½ · (1595 + 3657 + 3756 cm⁻¹) · h c Nₐ / 4184)
─────────────────────────
AE Dₑ = 232.23 kcal/mol  ≈ 232.5 kcal/mol (Curtiss G2/3 reference)

The Provenance notes on every emitted record include this distinction so it can’t get conflated in downstream comparisons:

notes = 'fetched from CCCBDB exp2x; atomization_energy_kcal_per_mol
is D_0 (0 K, ZPE-included); for D_e add ZPE = 0.5 * sum(
vibrational_fundamentals_cm_inv) converted to kcal/mol'

Programmatic equivalent + run on NIST geometry

From Python you can fetch the CCCBDB record, bridge its experimental geometry straight into a vibe-qc Molecule, and run an SCF on it. This fetches water, converts the NIST geometry, and runs RHF/STO-3G:

from vibeqc.fetch.references import (
    experimental_geometry_to_molecule_spec,
    fetch_cccbdb,
)
import vibeqc as vq

ANG_TO_BOHR = 1.0 / 0.529177210903

ref = fetch_cccbdb(cas="7732-18-5")                          # H₂O
spec = experimental_geometry_to_molecule_spec(ref, slug="h2o_nist")

mol = vq.Molecule([
    vq.Atom(a.z, [c * ANG_TO_BOHR for c in a.xyz_ang]) for a in spec.atoms
])
result = vq.run_job(mol, basis="sto-3g", method="rhf")
print(f"E(NIST geometry) = {result.energy:.8f} Ha")
# → -74.96302314 Ha (NIST r=0.958 Å vs Szabo-Ostlund r=1.0 Å: Δ ≈ +1 mHa)

Attaching CCCBDB columns to a regression-suite report

The regression-suite runner can fold the fetched NIST values into its report. Passing --include-experimental-reference cccbdb adds an experimental-reference column for each system in the run:

python -m examples.regression.run_suite \
    --systems h2o,ch4,hf --bases sto-3g --methods rhf,rks-lda \
    --include-experimental-reference cccbdb

The emitted summary.md gets an “Experimental references” section with the NIST values + DOI citation:

system

formula

AE D₀ (kcal/mol)

IE (eV)

μ (D)

vib fundamentals (cm⁻¹)

NIST source

h2o

H2O

219.35

12.621

1.857

1595, 3657, 3756

CCCBDB 7732-18-5

Cache & offline mode

Every successful fetch lands at $XDG_CACHE_HOME/vibeqc/fetch/<source_db>/<source_id>.json (default ~/.cache/vibeqc/fetch/). TTL is 30 days for OPTIMADE / MP / NOMAD / CCCBDB; infinite for COD because COD entries are immutable post-publish.

Two env-var levers:

VIBEQC_FETCH_CACHE_ROOT=/tmp/myrun  vqfetch canonical mgo_rocksalt
VIBEQC_FETCH_CACHE_ONLY=1           vqfetch reference --cas 7732-18-5

--cache-only (set the env var or pass the flag) makes vqfetch refuse to hit the network, useful for offline / CI / reproducibility.

Polite rate limiting (CCCBDB)

NIST is a US-government public resource shared by the whole community. client_cccbdb.py enforces:

  • ≥ 1 second between HTTP requests per process (per-process throttle).

  • Declared User-Agent identifying vibe-qc + version + repo URL so NIST can find us if there’s a problem.

  • Retry-After honoured on 429/503 with exponential backoff (5 s → 60 s, 3 attempts max).

The 30-day cache means re-fetches are rare in practice, and every cached hit is a request NIST doesn’t have to serve.

Where to read more

  • docs/user_guide/external_structures.md

    • full Phase-1 reference (CLI flags, every source, troubleshooting).

  • docs/user_guide/reference_data.md

    • full Phase-2 reference (CCCBDB schema, every property, regression integration).

  • docs/handover_structure_fetcher_smoke.sh / docs/handover_reference_data_fetcher_smoke.sh, the §12 / §11 acceptance harnesses (run them after a fresh install to verify the pipeline works on your machine).

  • docs/roadmap.md §§ v0.8.0 vqfetch external-data

    • v0.8.x vqfetch polish, what shipped in v0.8.0 and what’s queued in the maintenance line (runner_crystal.py for periodic parity, NIST WebBook fallback, ATcT integration, …).