Non-stationary point used in vib unit tests

[langner]

The coordinates we use for vibrational analysis, at least in some cases, is not a stationary point. While this probably does not influence the performance of parsers, it would be nice to do these calculations at startionary points. I'm not sure which programs check this and report it, but GAMESS-US does: https://github.com/cclib/cclib/blob/master/data/GAMESS/basicGAMESS-US2012/dvb_ir.out#L11303

(I dug this up from an old email on the ML)

[langner]

Kind of part of #45.

[berquist]

I think that this depends on whether or not a program projects out translations and rotations, just assumes the lowest six modes can be ignored and throws them out, or leaves them in. Q-Chem does the second and maybe the first (54 modes), GAMESS-US does the last (60 modes), Gaussian does the second, and ORCA does the second.

[langner]

I'd like to understand the details about this:

• What do you mean by "project out translation and rotations"?
• Does throwing out the first 6 modes make the vibrational analysis valid? Is a harmonic approximation used in QChem and all other programs (as GAMESS does)?
• Is this something you would recommend we address?

[ATenderholt]

ORCA prints the translations/rotations with the vibrations---the six lowest frequencies in dvb_ir and dvb_raman are all 0---but the parser ignores the first six modes and prints a message if any are non-zero. This is dangerous if it's a transition state structure and throws out the imaginary mode. It's also a problem if it's a linear molecule and should only be throwing out five modes. Maybe we should have something like hydrogen cyanide as a regression?

For Gaussian, there is a block of "Low frequencies" that we could parse since it includes the 6 low-frequency modes that it assigns as vibrations/translations.

As far as "projecting out transitions and rotations", I think this Gaussian page explains it: http://www.gaussian.com/g_whitepap/vib.htm

[berquist]

The general procedure is outlined here: http://www.gaussian.com/g_whitepap/vib.htm

Projection refers to the transformation from mass-weighted Cartesian coordinates to internal coordinates, and in the process of this transformation one can separate out translational and rotational modes from the vibrational modes. This can introduce some variation in the final results.

Compare the lowest 9 modes of the GAMESS-US and Gaussian outputs (from "Low frequencies"): -15.52 -8.02 -0.06 0.03 1.06 11.21 47.87 81.04 152.39 -6.0594 -4.5939 -0.0005 -0.0003 -0.0003 3.9500 52.7881 83.9570 148.1576

Gaussian doesn't complain about this; if it were the 5th or 6th mode, it would. I'm not sure that it's an issue.

A harmonic approximation is still used in Q-Chem.

Looks like Adam beat me to the punch.

[langner]

OK, so this means simply removing the lowest frequency rotational and translational modes. Still, the vibrational analysis is not valid if the geometry is not a stationary point, right? That is my main concern at this point.

I understand there are some subtlties here that can be considered. We can add some additional regressions to take them into account, of course.

[langner]

@berquist What do you think about this? Is there something specific action items we can identify here to make this concrete?

[berquist]

I'm confused because other than an old ORCA unit test that's probably a bad calculation or a bug, an ORCA regression, and two ADF unit tests, none of IR outputs present negative frequencies in the regular output (not talking about the "low frequencies" section because not all programs have that).

• The first step is to identify the problematic files. There is some inconsistency in which structure is used for each program (see https://github.com/cclib/cclib/issues/407).
• If a new geometry should be used, it can't be the same as the one used for geometry optimizations unless you're ok with a geo opt test only taking 1-2 steps.
• Programs do different things with symmetry and reorientation of inputs, which is a problem because the DFT grid is not invariant to this reorientation. The effect should be small for B3LYP but I don't have proof yet. Disabling reorientation is a possibility but this may affect testing vibsyms. I am against turning off lots of features selectively like this because it isn't the default usage, which is the most important case for us to cover; that should be for regressions.

A partial fix to the solution is

2. use that program's (+ version) converged geometry for a frequency calculation with all default symmetry enabled.

The most complete solution is to

0. switch to Hartree-Fock for most tests,
1. pick a uniform non-stationary starting geometry for most calculation types across all programs,
2. use that program's (+ version) converged geometry for a frequency calculation with all default symmetry enabled.

[langner]

I like the complete solution most - let's coordinate after v1.5.3 is out.