PubChemQC FCIDUMP build

Scripts that turn PubChemQC B3LYP/6-31G* records (Nakata and colleagues, CC BY 4.0, from MolSSI's public mirror on Hugging Face, molssiai-hub/pubchemqc-b3lyp) into CAS(8e,8o) FCIDUMP files with PySCF, for the BECOME energy catalog at https://become.belovedecosystem.com/energy/.

• scripts/pcq_prep.py streams one source file of a subset, keeps neutral singlets with at most maxorb orbitals (basis-function count minus heavy-atom count), applies the optional selection settings below, drops CIDs in an optional skip list, and with shards=auto sizes the shard count to the candidate count so dense files finish inside the per-job time budget.
• scripts/pcq_prep_group.py and scripts/pcq_pool.py are the pooled mode: many source files are scanned in grouped prep jobs, their candidates are pooled (repeated CIDs dropped) and re-sharded.
• scripts/pcq_convert.py rebuilds RHF in 6-31G*, checks it against PubChemQC, and writes the FCIDUMP.
• scripts/fcidump_check.py independently checks each FCIDUMP.
• scripts/pcq_run.py runs a batch in parallel and writes catalog.jsonl.
• scripts/pcq_merge.py merges the out/ directories of several shards into one tree (first shard wins on a duplicate CID; duplicates are listed in merge_duplicates.jsonl).
• scripts/pcq_finalize.py re-checks every file and writes the build library: library/catalog.jsonl and, for each molecule, library/published/<CID>/ with FCIDUMP_pcq<CID>_cas8e8o.gz (plus a full-space file for molecules with at most 60 orbitals), pcq<CID>.xyz, entry.json, SHA256SUMS for the uncompressed text and SHA256SUMS-gz for the .gz archives.
• scripts/pcq_package.py packages one or more build libraries into the deposited layout: the active-space files uncompressed as become-fcidump-pcq/fcidump/pcq<CID>_6-31g_cas8e8o.fcidump, the geometries as become-fcidump-pcq/xyz/pcq<CID>.xyz, the index catalog_index.csv (one row per entry), SHA256SUMS.txt and, with --tar, a reproducible become-fcidump-pcq.tar.gz.

The Build PubChemQC FCIDUMPs workflow runs prep and build across many GitHub runners at once, then a finalize job merges every shard and uploads the build library as the library artifact. The deposited catalog is made from the four sources' libraries with

python scripts/pcq_package.py <out> <chon300nosalt library> <chnopsfcl300nosalt library> \
    <chnopsfclnakmgca500 library> <b3lyp_pm6 library> --tar [--entry-numbers <CSV of cid,i>]

Column i of catalog_index.csv is the catalog entry number; without --entry-numbers the entries are numbered in packaging order.

Selection settings for the published catalog

The published catalog (125,999 molecules) comes from four PubChemQC sources. Each was built with the workflow inputs below; inputs not listed keep their defaults (limit=0, pool=0, minorb=0, no_multifrag=0, empty exclude_z, allow_z and skip). shards, per_shard, the pooled shard count and budget_min only spread the work over runners and do not change which molecules are selected.

Source	subset	file_from-file_to	maxorb	Other selection inputs
b3lyp_pm6_chon300nosalt	chon300nosalt	0-86	150	none
b3lyp_pm6_chnopsfcl300nosalt	chnopsfcl300nosalt	0-148	150	skip=data/skip_chnopsfcl300nosalt.txt.gz
b3lyp_pm6_chnopsfclnakmgca500	chnopsfclnakmgca500	0-346	150	pooled (pool=8), exclude_z=11,12,19,20 (Na, Mg, K, Ca), no_multifrag=1, skip=data/skip_chnopsfclnakmgca500.txt.gz
b3lyp_pm6, source file 1 (CIDs 253,697-507,790)	b3lyp_pm6	1-1	200	minorb=150, allow_z=1,5,6,7,8,9,14,15,16,17,35, no_multifrag=1, skip=data/skip_b3lyp_pm6.txt.gz, pooled (pool=16)

Each skip list holds the CIDs that earlier builds had already processed, so a molecule is built once. pcq_convert.py then applies the same gates to every source (neutral closed-shell singlet, elements H, B, C, N, O, F, Si, P, S, Cl, Br, no formal charges, PubChem identity kept after PM6, single fragment, B3LYP energy reproduced, stable RHF).

Method notes

The functional is named explicitly (b3lypg), degenerate orbitals are fixed by a fixed generic operator, a phase rule is applied to every orbital, gzip is written with mtime 0, every file is re-read and its integrals checked (no correlated energy is computed or stored), and molecules too small for an 8e8o active space are rejected. Repeat runs agree within about 1e-9 in every integral but are not byte-identical, so compare files by tolerance, not by hash.

License

The code in this repository is licensed under the Apache License 2.0 (see LICENSE). Copyright 2026 Beloved Ecosystem, Inc. The PubChemQC input data are not part of this repository; they are licensed separately under CC BY 4.0 by their authors.