seamm_thermochemistry

Shared atomic reference-energy database and formation-energy arithmetic for SEAMM.

Why

A raw total energy from Gaussian, ORCA, Psi4, or VASP has an arbitrary, code-dependent zero -- meaningless to a non-expert SEAMM user, and not comparable across codes. Re-expressing it as an energy/enthalpy of formation (relative to the elements in their standard states) fixes both problems at once. See the design doc for the full rationale: ~/Sites/reference-energy/2026-07-24_reference-energy/.

Today, gaussian_step, psi4_step, and vasp_step each carry an independent, mostly-duplicated copy of this logic and its data (a ~5000-column, mostly-empty CSV per molecular code; a separate workbook for VASP). This package is the single shared replacement:

• db.py -- a SQLite-backed ThermoDB with two tables (element: the experimental reference data every code needs; atom_energy: one row per computed atomic reference energy, keyed by element/code/method/ref_type/ settings, with room for provenance) and a small helper API (add_element, add_atom_energy, get_reference_energies, missing, dump_*_csv, ...). Zero third-party dependencies beyond seamm_util.
• formation.py -- atomization_energy() and formation_energy(), the arithmetic every plugin's calculate_enthalpy_of_formation currently reimplements, generalized to also produce a ZPE-free energy of formation when no harmonic thermochemistry has been run.
• importers.py -- one-off loaders from the three legacy master files (Paul's experimental-data workbook, the VASP element-energy workbook, and the gaussian_step wide CSV) into a ThermoDB. Needs the import extra (pandas, openpyxl). (psi4_step's CSV is not imported: it is a copy of gaussian_step's, i.e. Gaussian numbers, not Psi4 results.)

Two reference conventions, one schema

ref_type on atom_energy distinguishes:

• "atom" -- isolated gas-phase atom (Gaussian/Psi4/ORCA's convention, and the target for VASP once the atom-in-a-box calcs are wired in). Pairs with an experimental anchor (element.dfH0_0K / dfH0_298K) to give a true, cross-code-comparable energy/enthalpy of formation.
• "element_phase" -- energy per atom of the element's standard-state phase (bulk metal, graphite, O2(g), ...). VASP's existing convention (element_energies.csv's plain <method>@<encut> columns) -- no experimental anchor needed, and useful as a fallback reference for elements (e.g. Mn) where the free atom is a poor DFT target.

formation_energy(..., anchor=True) matches the existing gaussian_step/psi4_step enthalpy-of-formation arithmetic exactly (and gives an energy of formation, not enthalpy, when system_energy excludes ZPE and the 0 K anchor is used). formation_energy(..., anchor=False) matches vasp_step's existing DfE0 exactly. Both are exercised in tests/test_formation.py against hand-worked numbers.

Status

The reference database is published on Zenodo (a DOI per version) and fetched with seamm-thermochemistry-installer install -- not bundled in the Python package. gaussian_step's calculate_energy_of_formation already consumes it in production. Current coverage: the full Gaussian composite-method/basis grid, VASP (PBE family, both the isolated-atom and standard-state-phase conventions), and ORCA (several DFT methods across the full def2 basis family), all vetted and imported via seamm-thermochemistry-import-orca / the importers module. There are no Psi4 atom energies yet: the earlier "psi4" rows were copies of Gaussian's.

Computing ORCA atom energies

scripts/orca_atom_multistart.py computes the ORCA atomic reference energies. An atom's reference energy for a method is defined as the lowest-energy SCF solution of that method at the experimental ground-state spin multiplicity. That is well defined even where the method orders the atom's states differently from experiment. There, though, the atomization route to formation energies is itself becoming unreliable, and for careful work reaction energies to well-known species (e.g. H2O -> H2 + 1/2 O2) are the better tool. Formation energies remain far more meaningful than raw total energies, for users and for machine-learning training data alike.

A single SCF start is not enough to find that solution for open-shell atoms: Fe from ORCA's PModel guess lands in 3d7 4s1, 120 kJ/mol above 3d6 4s2, and small setting changes flip Nb and Ho between solutions over 80 kJ/mol apart. So the script runs several starts per element, method and basis (PBE0/def2-SV(P) orbitals, PModel, HCore, Hueckel, PAtom; from La on, where ORCA has no Hueckel or PAtom guess, Fermi-smeared PBE orbitals instead), shares every state found in one basis with all the others, and keeps the lowest SCF energy. <S**2> is recorded and contamination over 20% flagged, but it does not disqualify a solution. It uses exact exchange (NoCOSX), because ORCA's default COSX gets some lone atoms wrong (He ~1, Na and Mg ~5 kJ/mol), and full orbital convergence (ConvCheckMode 0) for the MP2 part of double hybrids. It writes a Results.csv that seamm-thermochemistry-import-orca reads, plus a log of every run and of each choice.

ORCA works in a scratch directory under $TMPDIR -- node-local on a cluster -- and only orca.out and orca.gbw are kept in the run directories. With hundreds of runs at once, ORCA's temporary files on a shared filesystem made the runs I/O-bound and many times slower. A run always starts from its own guess: orbitals left by an earlier, unfinished attempt are not picked up.

One known simplification: the settings column is a single free-form string (e.g. "encut=700eV") rather than normalized basis/cutoff columns -- fine across today's producers, may want normalizing if that stops being true.

Quick start

from seamm_thermochemistry import ThermoDB, formation_energy

with ThermoDB("my_reference.db") as db:
    db.add_element(1, "H", dfH0_0K=216.034, standard_state="1/2 H2(g)")
    db.add_element(8, "O", dfH0_0K=246.79, standard_state="1/2 O2(g)")
    db.add_atom_energy("H", "gaussian", "CBS-QB3", -1312.0)
    db.add_atom_energy("O", "gaussian", "CBS-QB3", -197400.0)

    dfE = formation_energy({"H": 2, "O": 1}, system_energy, db, "gaussian", "CBS-QB3")

Copyright

Copyright (c) 2026, MolSSI SEAMM