Segmentation faults + symmetry issues

thorium · February 8, 2018, 7:34pm

Hi,

I am not sure if the build category is the right place to post this, or if I should post it on GitHub or here. Apologies if this is the wrong place. I am using psi4 1.1add49b9 on a Linux Redhat system, and have installed psi4 within a conda environment. I have two different issues I wanted to ask about:

I get a segmentation fault for the following input:

import psi4
import numpy as np

molecule h_lat {
A = 2.0000
#atom x y z
#left top
H -A A 0.000
#middle top
H 0.000 A 0.000
#right top
H A A 0.000
#left middle
H -A 0.000 0.000
#origin
H 0.000 0.000 0.000
#right middle
H A 0.000 0.000
#left bottom
H -A -A 0.000
#middle bottom
H 0.000 -A 0.000
#right bottom
#H A -A 0.000
symmetry c1
}

psi4.set_options({‘reference’: ‘rhf’, ‘scf_type’: ‘df’, ‘maxiter’: 500, ‘stability_analysis’: ‘follow’})

scf_e, scf_wfn = psi4.energy(‘scf/sto-3g’, molecule=h_lat, return_wfn=True)

Why does this happen? I know the geometry is unusual, but the segmentation fault occurs for a variety of values of A.

Given any input molecule, when I specify symmetry in the molecule geometry, scf_wfn.H() returns the actual Hamiltonian. But when I don’t specify the symmetry, I get the memory location of the Hamiltonian.

import psi4
import numpy as np

molecule h2 {
H 0.0 0.0 0.0
H 0.0 0.0 0.7
}

psi4.set_options({'reference': 'rhf', 'scf_type': 'df'})

scf_e, scf_wfn = psi4.energy('scf/sto-3g', molecule=h2, return_wfn=True)
mints = psi4.core.MintsHelper(scf_wfn.basisset())

h=np.asarray(scf_wfn.H())
print(h)

gives output: <psi4.core.Matrix object at 0x2b67243bc470>. But when I add symmetry c1 in the geometry, I get:

[[-1.14765024 -1.00692423] [-1.00692423 -1.14765024]]
as my output. Is this intentional?

Thank you, and please let me know if I should post this question elsewhere.

dgasmith · February 8, 2018, 7:43pm

can you post the geometry?
Wavefunctions with symmetry cannot be represented by a single numpy array. So when using the NumPy casting technology nothing happens due its default pass through. Please see the Irrep section here: http://www.psicode.org/psi4manual/master/numpy.html

thorium · February 8, 2018, 7:49pm

I have posted the geometry for both issues. Is there some other geometry you want that I’m missing?
So for wfns with symmetry, I should create numpy matrix/vector objects first, and then pass them to psi4 and then run scf on them? Am I understanding this correctly? Is this also why in the instances I don’t get a segmentation error, I’m getting a divide-by-zero error originating from vecutil.py?

Thanks!

dgasmith · February 8, 2018, 8:06pm

For 1) what does Psi4 show as the XYZ? Im not at a computer at the moment to try this myself.

For 2) what you should do is write scf_wfn.H().to_array() to get back out the H matrix with irreps (a list of NumPy arrays). Which operations are you doing when you get seg faults?

thorium · February 9, 2018, 7:13pm

Here is the psi4 geometry, sorry for the delay in sending this:

Molecular point group: c1
    Full point group: Cs

    Geometry (in Angstrom), charge = 0, multiplicity = 1:

       Center              X                  Y                   Z               Mass
    ------------   -----------------  -----------------  -----------------  -----------------
           H          0.000000000000     2.474873734153     0.000000000000     1.007825032070
           H         -1.414213562373     1.060660171780     0.000000000000     1.007825032070
           H          0.000000000000    -0.353553390593     0.000000000000     1.007825032070
           H          0.000000000000     1.060660171780     0.000000000000     1.007825032070
           H          0.000000000000    -0.353553390593     0.000000000000     1.007825032070
           H          0.000000000000    -1.767766952966     0.000000000000     1.007825032070
           H          0.000000000000    -0.353553390593     0.000000000000     1.007825032070
           H          1.414213562373    -1.767766952966     0.000000000000     1.007825032070

  Running in c1 symmetry.

The segmentation fault occurs for energy(‘scf’) with a few different reference/basis set/scf_type options.

loriab · February 12, 2018, 4:14pm

Are your atoms 3, 5, 7 at the same point in space? I’m surprised that molecule got through.

Small consolation, but it looks like my new mol parser fares better:

    units Angstrom
    0 1
    --
    0 1
    H                            -A                  A     2.474873734153
    H                1.414213562373                  A     1.060660171780
    H                             A                  A    -0.353553390593
    H                            -A     0.000000000000     1.060660171780
    H                0.000000000000     0.000000000000    -0.353553390593
    H                             A     0.000000000000    -1.767766952966
    H                            -A                 -A    -0.353553390593
    H               -1.414213562373                 -A    -1.767766952966

    A         =   -0.0000000000


    Molecular point group: c1
    Full point group: C2v

    Geometry (in Angstrom), charge = 0, multiplicity = 1:

       Center              X                  Y                   Z       
    ------------   -----------------  -----------------  -----------------
           H          0.000000000000     0.000000000000     2.474873734153
           H          1.414213562373     0.000000000000     1.060660171780
           H          2.828427124746     0.000000000000    -0.353553390593
           H         -1.414213562373     0.000000000000     1.060660171780
           H          0.000000000000     0.000000000000    -0.353553390593
           H          1.414213562373     0.000000000000    -1.767766952966
           H         -2.828427124746     0.000000000000    -0.353553390593
           H         -1.414213562373     0.000000000000    -1.767766952966

thorium · February 12, 2018, 10:34pm

So given that in my initial input geometry atoms 3, 5, and 7 don’t have the same coordinates (given that A != 0), why does the molecule parser make them the same? And how can I fix this?

Also, just in general, I am curious to know how the geometry is parsed. Is the user-input geometry not used as given?

loriab · February 15, 2018, 5:39pm

It gets parsed here. But given that I just removed that fn from the dev psi a couple days ago, I wouldn’t examine it too closely.

Whether geom is used as given depends on Cartesian (yes; subject to com/orientation) vs. ZMat/variables (gets complicated). As a stopgap, feed psi non-variable Cartesians. Sorry about that, but I’m not inclined to debug code that’s soon to be gone. Certainly you’re in the right and psi is in the wrong molecule-parsing-wise.