IR Intensities v2.0

[kcpearce]

Description

Will add reference implementation for computing IR intensities of molecules.

What are your new additions? Please provide a brief list.

• New Features
  • [x] Psi4NumPy reference implementation using Andy's dipole derivatives

Status

• [x] Click when ready for review-and-merge

[dgasmith]

Thanks for tackling this!

[kcpearce]

@dgasmith @loriab I'm currently testing the frequencies against those computed by Psi4. Does the wfn variable store the IR intensities as well to compare against? If so, what are they called?

[loriab]

Should be in a field attached to wfn.frequency_analysis . let me know if it's missing. https://github.com/psi4/psi4/blob/master/tests/freq-isotope2/input.dat#L101-L116

[kcpearce]

Stupid question: wfn.frequency_analysis()['IR_Intensity'] contains

----------------------------------------
        Datum infrared intensity        
----------------------------------------
Data:     [  0.             0.             0.             0.             0.             0.             2.7682201519 163.8780634213   0.             0.             0.             0.            16.5538835573
   0.            27.5122695971   0.             0.            63.3981457698]
Units:    [km/mol]
doi:      None
Comment:  
Glossary: 
----------------------------------------

How do I access the 'Data' member?

[loriab]

ah yes, it's a Datum (https://github.com/MolSSI/QCElemental/blob/master/qcelemental/datum.py#L12). get it with .data, I think.

[kcpearce]

Ready for review!

[kcpearce]

Looks like I'm failing quite a few tests in Travis py37. Any ideas as to why?

[loriab]

ah, that can happen when run infrequently.

• the mp3 and mp2.5 cases may be b/c default changed from conv to df. could you try adding set mp_type conv in those?
• some others have to do with the L2 switch: https://travis-ci.org/github/psi4/psi4numpy/jobs/748406893#L1774 . needs investigating to see if API changed or used a little wrong here. braket type is like 4-center xx_xx or 3-center xs_xx or 2-center xs_xs

[dgasmith]

Looks like on Python 3.7 Psi4 1.4a3.dev51+83a8459 is failing, but 1.3.2+ecbda83 passes fine.

----------------------------- Captured stderr call -----------------------------
1388Traceback (most recent call last):
1389  File "/home/travis/build/psi4/psi4numpy/Coupled-Electron-Pair-Approximation/DF-LCCD.py", line 39, in <module>
1390    R, F = integrals_DF(mol)
1391  File "/home/travis/build/psi4/psi4numpy/Coupled-Electron-Pair-Approximation/integrals.py", line 80, in integrals_DF
1392    Ppq = np.squeeze(mints.ao_eri(zero_bas, aux, basis, basis))
1393RuntimeError: 
1394Fatal Error: Bad BraKet type in Libint2TwoElectronInt
1395Error occurred in file: /scratch/psilocaluser/conda-builds/psi4-multiout_1608334598638/work/psi4/src/psi4/libmints/eri.cc on line: 198
1396The most recent 5 function calls were:
1397
1398psi::IntegralFactory::eri(int, bool, bool)
1399psi::MintsHelper::ao_eri(std::shared_ptr<psi::BasisSet>, std::shared_ptr<psi::BasisSet>, std::shared_ptr<psi::BasisSet>, std::shared_ptr<psi::BasisSet>)

A little surprising since I'm pretty sure we build DF integrals in Psi4 tests. Any ideas @andysim @loriab.

[andysim]

Sorry, that's a Libint2 constraint. (Ax|mn) is allowed, but not (xA|mn), so it's a trivial fix. Here x is the zero basis set, and A is the aux.

[dgasmith]

Thanks @andysim. @kcpearce I know its not your PR, but would you mind patching?

[kcpearce]

No problem. @andysim I assume the analogous (mn|Ax) is allowed, but not (mn|xA)?

[andysim]

Correct! I just made PR to your branch after I found two offending integral calls in the CEPA code. If there are any calls like you mentioned in your previous comment, they'll need to be patched also. Please let me know if there are any more problems - I was hoping not to break too much with the L2 upgrade, but there seem to be a few small things.

[kcpearce]

@andysim I think I've caught all of the offending calls in this last commit including those two in the CEPA code

[andysim]

Awesome - thanks so much for fixing those. It never crossed my mind to check this before

[kcpearce]

1g test is still failing convergence for with a 1e-10 tolerance. Should I loosen it?

  Memory set to 476.837 MiB by Python driver.
    Point group...........................................................................PASSED
    Nuclear repulsion energy..............................................................PASSED
    SCF Energy............................................................................FAILED
----------------------------- Captured stderr call -----------------------------
[NbConvertApp] Converting notebook /home/travis/build/psi4/psi4numpy/Tutorials/01_Psi4NumPy-Basics/1g_basis-sets.ipynb to script
[NbConvertApp] Writing 5720 bytes to /tmp/tmp4uug5474/1g_basis-sets.py
Traceback (most recent call last):
  File "/tmp/tmp4uug5474/1g_basis-sets.py", line 115, in <module>
    psi4.compare_values(refscf, eb, 10, "SCF Energy")                                  
  File "/home/travis/miniconda/envs/p4env/lib/python3.7/site-packages/psi4/driver/qcdb/testing.py", line 104, in _mergedapis_compare_values
    return qcel.testing.compare_values(expected, computed, **kwargs)
  File "/home/travis/miniconda/envs/p4env/lib/python3.7/site-packages/qcelemental/testing.py", line 171, in compare_values
    return return_handler(allclose, label, message, return_message, quiet)
  File "/home/travis/miniconda/envs/p4env/lib/python3.7/site-packages/psi4/driver/p4util/testing.py", line 189, in _psi4_true_raise_handler
    raise TestComparisonError(message)
psi4.driver.p4util.exceptions.TestComparisonError:  SCF Energy: computed value (-228.957630058814) does not match (-228.957630059008) to atol=1e-10 by difference (0.000000000194).

[andysim]

Yeah, that's being tested a bit too tight. You can set the integral tolerance lower, but I think just loosening the check makes more sense. The answer is nondeterministic to some degree when running in parallel anyway, so super tight checks don't really make sense

[kcpearce]

Lowered the tolerance to 7, and looks like we're passing Travis. Finally ready for merge!

[dgasmith]

Nice!