Bad wavefunction with spin polarized SPO using LCAO

[shivupa]

After a discussion with Paul via email, I'm opening an issue on the following problem we've been having.

Relevant files are in the following gist.

The systems for which we have encountered problems are dipole-bound anions. These are systems in which an electron binds in the dipole field of the molecule. They have very extended charge distributions. We have been able to calculate electron binding energies for several of these systems using Casino, but attempts to do so with QMCPACK have failed. The most recent such system that we have attempted is CH3CN.

The neutral molecule presents no problems: the problems are with the anion.
The calculations are all-electron with 1-body and 2-body Jastrow factors. Our most recent calculations used the 3.5.0 release of QMCPACK. These were all-electron calculations with cusp corrected orbitals using the standard cusp correction algorithm. The VMC without a Jastrow (but with cusp corrected orbitals) does agree with HF (-131.97 ± 0.19 for VMC and -131.939 for HF).

We first optimized the Jastrow parameters for the neutral using cutoffs optimized in Casino and then used these as starting parameters for the anion. However, the VMC optimization on the anion failed with variances ~ 10^14. The Jastrow factors used splines and were optimized using the OneShiftOnly optimizer.

Pure variance (both unreweighted and reweighted) minimization behaved similarly. In both optimization approaches the high variances begin from the first iteration.

I attached plots of the starting Jastrow factors. They were generated using a short optimization on the neutral. It seems like a reasonable starting guess.

We've tried another dipole-bound anion system, C3H2 and were similarly unsuccessful. In that system there is the possibility of collapsing to a valence bound anion which could cause complications so we switched to CH3CN. Similarly, a long time ago we tried a model water system which was also unstable.

The binding of the systems we are trying to treat is very diffuse compared to common anionic systems, but I am unsure if this would cause any complications for QMC.

Let me know if I can provide any other relevant information.

Thanks,
Shiv

[ye-luo]

I ran your input with current develop. It seems that once the extra electron is added, the inital VMC already goes wrong. I tried without Jastrow with AoS and SoA code. The AoS code already gives 10^14 variance while the SoA code shows very absurd energy trajectory, having some jump like a step function, although the energy values are around -130.
v3.1.1, v3.4.0 and v3.5.0 AoS builds both show insane huge variance. I think something is wrong with the orbital or cusp correction.

[prckent]

@shivupa Thanks for reporting the problem.

[anbenali]

I am looking at the problem now. will update this afternoon.

…

On Mon, Nov 5, 2018 at 8:02 AM Paul R. C. Kent ***@***.***> wrote: @shivupa <https://github.com/shivupa> Thanks for reporting the problem. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1145 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AXl6w2XgQJLqVJCnTJB2uph5Vv3kK27Fks5usEUHgaJpZM4YNe6L> .

-- ----------------------- Anouar Benali, PhD Leadership Computing Facility Argonne National Laboratory Building 240 Office - 2127 9700 S Cass Av., Argonne Il, 60439 (630) 252-0058

[shivupa]

Thanks I added the gamess input file to the gist above if needed.

[anbenali]

Hi Shiv, I am norowwing down the issue but I would like to go through the whole process and for that I need your gamess input file. Could you please send it to me? I could not find which basis set you used for Carbon 1. Thanks Anouar

…

On Mon, Nov 5, 2018 at 11:09 AM Anouar Benali ***@***.***> wrote: I am looking at the problem now. will update this afternoon. On Mon, Nov 5, 2018 at 8:02 AM Paul R. C. Kent ***@***.***> wrote: > @shivupa <https://github.com/shivupa> Thanks for reporting the problem. > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#1145 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AXl6w2XgQJLqVJCnTJB2uph5Vv3kK27Fks5usEUHgaJpZM4YNe6L> > . > -- ----------------------- Anouar Benali, PhD Leadership Computing Facility Argonne National Laboratory Building 240 Office - 2127 9700 S Cass Av., Argonne Il, 60439 (630) 252-0058

-- ----------------------- Anouar Benali, PhD Leadership Computing Facility Argonne National Laboratory Building 240 Office - 2127 9700 S Cass Av., Argonne Il, 60439 (630) 252-0058

[shivupa]

I managed to modify the converter on this branch. The converter now takes the labels from the gamess output and this fixes any issues with needing to modify the files manually. However, I still experienced the extremely high variances. Hopefully this helps narrow it down.

[prckent]

Thanks. It indeed seems that there is at least a problem with the converter although we can't also yet rule out the cusp correction. Results have not changed in years, so I think it is most probable that this is a legacy problem with the converter and limited non-error trapped capabilities. When the new cusp SoA implementation (#1150) is done we will be able to cross-check with PySCF, which uses a different HDF5 based route.

I would not use a wavefunction with high variance.

[anbenali]

Dear Shiv, Sorry for the delay. It took a while to eliminate all sources of error, but it seems like the converter from gamess is probably busted for this kind of calulations.. while gamess was a great code in he past, fixing the converter for this problem will take sometime... instead using pyscf might be a better solution as it scales significantly more than gamess and as you can see in the results in the box file, leads to a stable behavior. Please let me know if you need any assistance using pyscf. Anouar https://anl.box.com/s/mmpb7fxlhvkp0ympghiebhlduplnxhc0

…

On Sun, Nov 11, 2018 at 11:22 AM Paul R. C. Kent ***@***.***> wrote: Thanks. It indeed seems that there is at least a problem with the converter although we can't also yet rule out the cusp correction. Results have not changed in years, so I think it is most probable that this is a legacy problem with the converter and limited non-error trapped capabilities. When the new cusp SoA implementation (#1150 <#1150>) is done we will be able to cross-check with PySCF, which uses a different HDF5 based route. I would not use a wavefunction with high variance. — You are receiving this because you were assigned. Reply to this email directly, view it on GitHub <#1145 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AXl6wwSIfy5MjEnNZ02CzMpIo1dTU_Roks5uuFzigaJpZM4YNe6L> .

-- ----------------------- Anouar Benali, PhD Leadership Computing Facility Argonne National Laboratory Building 240 Office - 2127 9700 S Cass Av., Argonne Il, 60439 (630) 252-0058

[anbenali]

Dear Shiv, Unfortunately, After adding the diffuse functions on the C- atoms, calculations failed. At least now I am convince that the issue cannot be related to a bad converter issue. I will investigate further and keep you updated Thanks Anouar

…

On Wed, Nov 28, 2018 at 7:21 PM Anouar Benali ***@***.***> wrote: Dear Shiv, Sorry for the delay. It took a while to eliminate all sources of error, but it seems like the converter from gamess is probably busted for this kind of calulations.. while gamess was a great code in he past, fixing the converter for this problem will take sometime... instead using pyscf might be a better solution as it scales significantly more than gamess and as you can see in the results in the box file, leads to a stable behavior. Please let me know if you need any assistance using pyscf. Anouar https://anl.box.com/s/mmpb7fxlhvkp0ympghiebhlduplnxhc0 On Sun, Nov 11, 2018 at 11:22 AM Paul R. C. Kent ***@***.***> wrote: > Thanks. It indeed seems that there is at least a problem with the > converter although we can't also yet rule out the cusp correction. Results > have not changed in years, so I think it is most probable that this is a > legacy problem with the converter and limited non-error trapped > capabilities. When the new cusp SoA implementation (#1150 > <#1150>) is done we will be > able to cross-check with PySCF, which uses a different HDF5 based route. > > I would not use a wavefunction with high variance. > > — > You are receiving this because you were assigned. > Reply to this email directly, view it on GitHub > <#1145 (comment)>, > or mute the thread > <https://github.com/notifications/unsubscribe-auth/AXl6wwSIfy5MjEnNZ02CzMpIo1dTU_Roks5uuFzigaJpZM4YNe6L> > . > -- ----------------------- Anouar Benali, PhD Leadership Computing Facility Argonne National Laboratory Building 240 Office - 2127 9700 S Cass Av., Argonne Il, 60439 (630) 252-0058

-- ----------------------- Anouar Benali, PhD Leadership Computing Facility Argonne National Laboratory Building 240 Office - 2127 9700 S Cass Av., Argonne Il, 60439 (630) 252-0058

[shivupa]

Anouar,

I appreciate the help and the updates!

Thank you,
Shiv