Repository navigation
Feature: use a completed job's final MOs as the initial guess for a follow-on calculation #75
Description
Cross-code feature request from Andy (not specific to ORCA or any single code) — capture it now while it's fresh, design/implementation deferred.
When starting a new calculation from an existing completed one (e.g. duplicating a job to run a follow-on calculation, or restarting after changing theory/basis), it should be possible to ask ECCE to read the final converged molecular orbitals from the prior calculation's output and use them as the initial guess ("read wavefunction from previous job" / restart-from-MOs), rather than always starting from a generic initial guess (core Hamiltonian, atomic superposition, etc.).
Each code that supports this natively has its own mechanism and file format:
- NWChem:
vectors input <file>/ MOs stored in a.movecsfile. - Gaussian:
Guess=Readwith a checkpoint (.chk) file carried over. - ORCA:
%moinp "<file>.gbw"pointing at a prior.gbwfile, or the simpler! MOREADkeyword.
Needs, at minimum:
- A UI affordance in the calc editor / theory dialog for "use MOs from calculation X" (or "use MOs from the calculation this was duplicated from"), probably per-code since the underlying mechanism differs.
- The relevant orbital file to be retained/copied alongside a job's other outputs (currently these binary MO files, where they exist at all per code, may not be treated as first-class retained outputs the same way
.gbw/.movecs/.chkwould need to be). - Each code's
ai.<code>input generator to emit the correct restart directive when this is requested, pointing at the correct local copy of the prior job's MO file.
No design decisions made yet on the UI or the cross-code abstraction (if any) -- this issue is just to record the ask before it's forgotten, for whoever picks it up.
Activity
ohlincha commentedon Sep 22, 2026
Investigated 2026-09-22. Recording what the work actually is, since it is
larger than it looks and spans three layers.
Current state
No generator supports reading MOs. Zero references to
vectors input/movecsinai.nwchem,guess=read/oldchkinai.gauss16, orMORead/moinpinai.orca. NWChem's Initial Guess dropdown offers only"NWChem Atomic Guess" and "Core Hamiltonian".
The files are partly available already.
<DataFiles>declaresauxiliary MO outputs for NWChem (
movec.nw_mo) and Gaussian 16(
fort.7), so those are uploaded to the data server when a job finishes.ORCA declares none — its
.gbwwould have to be added.The duplicate machinery is clean and extensible.
CalcMgr::OnDuplicate2Clickisduplicate()followed byresetWithLastGeom(dup), andresetWithLastGeomis a self-containedfunction that rewrites the duplicate's fragment from GEOMTRACE. A third
variant would follow the same shape.
So the work is four pieces
and "Duplicate Setup with Last Geometry", per the maintainer's
suggestion. Mechanically the easiest part.
output into the duplicate, and record where it landed. This is the
part with no precedent:
resetWithLastGeomonly moves numbers thatare already in the calculation's properties, whereas this moves a
file between calculations on the data server.
NWChem
vectors input <file>; Gaussianguess=readwith thecheckpoint carried over (note ECCE currently declares
fort.7, thetext MO dump, not a
.chk— so Gaussian may need a differentauxiliary file than the one declared); ORCA
! MOReadplus%moinp "<file>.gbw", with the.gbwadded to<DataFiles>first.must be offered only when the duplicate actually has an MO file, or
it becomes another silently-dropped setting.
Suggestion
Do NWChem first and alone. It is the only code where the auxiliary file
declared is the one actually needed,
nwchemis installed here so thewhole chain can be verified by running it, and it proves or disproves
piece 2 — which is the one with no existing pattern to copy. Gaussian and
ORCA then follow with their own file handling.
Worth checking early whether the uploaded
movec.nw_mois genuinelyreusable: NWChem's MO vectors are basis-set dependent, so the follow-on
calculation has to use the same basis or the read fails. That constraint
should probably be enforced in the UI rather than discovered at run time.
ohlincha commentedon Oct 9, 2026
Extension: basis set projection. The follow-on calculation may use a different (usually larger) basis than the one the MOs come from. The guess is then the old orbitals projected onto the new basis, a standard way to converge large-basis or difficult cases from a cheap small-basis run.
The main use is unrestricted DFT on large systems that are hard to converge: open-shell transition-metal complexes, radicals, broken-symmetry states. The small-basis run settles the spin state and occupation pattern cheaply, and the projection carries both into the large basis, both α and β sets. Requirements:
What each code offers natively:
vectors input project <old basis name> <file.movecs>. The old basis must be defined in the same input under that name, so the generator also writes the old basis block.! MOReadwith%moinp "old.gbw". ORCA projects automatically when the basis differs (%scf GuessMode FMatrix|CMatrix end).Guess=Readwith%OldChk=old.chk. It projects automatically when the basis differs.UI: in the Calculation Editor, "Initial guess: MOs from ". When that calculation's basis differs from the current one, show "projected from ", so the user knows projection is used.
ORCA: Duplicate with Final Orbitals starts the copy from the source's…
ohlincha commentedon Oct 15, 2026
Stage 1 (ORCA only) is on branch
wip/75-moreadand is not merged. It is planned for 9.2..gbwis kept in its outputs..gbwasguess.gbwin.! MOReadand%moinp "guess.gbwin".tests/launch/orca_guess_test.py).Correction to the investigation above: declaring a file in
<DataFiles>does not upload it. The job monitor sends only the parse and output files plusfilesToTransfer. NWChem'smovec.nw_mois therefore probably not kept either. ORCA now opts in with<RetainAuxiliaryOutput>.Not yet verified:
Still to do: NWChem, Gaussian 16 and ECCE-QM.