Choose the wavefunction optimiser¶
Goal¶
Pick the CPMD wavefunction optimiser for a session that evaluates many nearby geometries, such as a minimisation, a saddle search, or a nudged elastic band. Every call after the first starts from the orbitals of the previous call, so the optimiser that suits a cold start from atomic orbitals is not always the cheapest for the warm calls that follow.
Measured on the Si3N4 cluster¶
One minimisation of a 7-atom Si3N4 cluster, BLYP at 70 Ry in a 14 Angstrom
cube with Kleinman-Bylander pseudopotentials, library route on 16 MPI ranks,
orbital convergence 1e-5.
The two runs differ only in the optimiser line of the &CPMD section.
Optimiser |
Wall time |
Force calls |
SCF steps |
Seconds per SCF step |
Minimum (eV) |
|---|---|---|---|---|---|
|
734 s |
33 |
515 |
1.26 |
-1396.271181 |
|
255 s |
23 |
296 |
0.64 |
-1396.271246 |
Both reach the same minimum: the two energies differ by 0.065 meV, inside the
orbital convergence tolerance.
ODIIS (orbital direct inversion in the iterative subspace) needed fewer and
cheaper SCF steps.
PCG (preconditioned conjugate gradient) with MINIMIZE runs a line
minimisation that costs about two energy evaluations per step.
Set ODIIS in the message¶
( cpmd = ( optimizeWavefunction = true,
convergenceOrbitals = 1.0e-5,
maxIter = 400,
odiisVectors = 10 ) )
odiisVectors = 10 renders ODIIS followed by 10 on the next line.
With no optimiser keyword at all, OpenCPMD 4.3 also uses preconditioned DIIS
with 10 vectors; its output then reads WAVEFUNCTION OPTIMIZATION BY PRECONDITIONED DIIS and MAXIMUM NUMBER OF VECTORS RETAINED FOR DIIS: 10.
Writing odiisVectors keeps the choice visible in the message.
odiis = true renders a bare ODIIS with CPMD’s default vector count, and
odiisOptions adds inline options such as NOPREC.
To keep the deck’s PCG line instead, use a directive:
( cpmd = ( optimizeWavefunction = true,
convergenceOrbitals = 1.0e-5,
maxIter = 400,
directives = [ ( keyword = "PCG MINIMIZE", args = [] ) ] ) )
Set only one optimiser in a message.
Check the rendered &CPMD section with CPMDC_DECK_OUT (see
debugging a deck).
When to keep PCG¶
cpmdc applies one deck to every call of a session, so the optimiser chosen
for the warm calls also runs the cold first call from atomic orbitals.
Keep PCG MINIMIZE for a system whose ODIIS calls stop at MAXITER: each
such call is reported as a failure (see troubleshooting),
and the optimiser choice does not change the converged state, only the path
to it.
Always set MAXITER¶
Without maxIter, the OpenCPMD path inserts MAXITER 40 into the first deck.
A warm call on a geometry that moved far from the previous one can need more
steps than that, and an SCF that stops at the limit fails the call.
Strict PCG line minimization¶
For strict orbital targets, apply tools/opencpmd_pcg_step.patch to the
OpenCPMD source linked into the native archive. The directional slope and
curvature both scale with the square of the orbital residual. An absolute
curvature floor of 1e-15 can suppress the interpolated step even when the
ratio of derivatives is well resolved. The patch divides both derivatives
by their largest magnitude before finding the root; equal derivatives
retain the supplied trial step. It does not change the convergence target.
cd /path/to/OpenCPMD
patch -p1 < /path/to/cpmdc/tools/opencpmd_pcg_step.patch
/path/to/cpmdc/tools/check_opencpmd_pcg_step.sh "$PWD" "$PWD/build/pcg-step-test"
/path/to/cpmdc/tools/rebuild_opencpmd_pcg_step.sh /path/to/cpmd-root
Run compilation inside a compute allocation. The check extracts the patched native function and calls it from C++ for 726 roots with derivative scales from 1e-300 to 1e300, plus degenerate and extrapolation cases. Expected output:
726 scaled native line-search roots and four degenerate/extrapolation cases pass
Relink libcpmdc against the rebuilt archive. Preserve the separate
opencpmd_omp_info.patch correction for calculator groups spanning nodes.
The algebraic check establishes scale invariance; native force convergence
must also be checked on the requested system and orbital target.