wying
January 5, 2023, 8:41am
1
I find it is hard to converge for the geometry optimization for most cation and anion pairs even with 500 iterations. I have met this issues for most of the ionic liquids. This is one example smile structure “CC[N+]1(C)CCCCC1.[O-]S(=O)(C(F)(F)C(F)F)=O”.
I am using the rdkit and psikit. I have tried many ways to tune the options as shown in the code. Can anyone help me with this issue?
sml = “CC[N+]1(C)CCCCC1.[O-]S(=O)(C(F)(F)C(F)F)=O”
def optimize(self, basis_sets= "scf/6-311pg**", return_wfn=True, name=None, multiplicity=1, maxiter=500,
opt_coordinates = "cartesian"):
if not name:
name = uuid.uuid4().hex
self.psi4.core.IO.set_default_namespace(name)
self.geometry(multiplicity=multiplicity)
self.psi4.set_options({'GEOM_MAXITER':maxiter,
'opt_coordinates': opt_coordinates,
#'BASIS_GUESS': True,
#'DAMPING_CONVERGENCE': 60,
#'D_CONVERGENCE': 0.000001,
#'E_CONVERGENCE': 0.000001
})
try:
#Geometry optimization
print("The basis set is ", basis_sets)
scf_energy, wfn = self.psi4.optimize(basis_sets, return_wfn=return_wfn)
self.wfn = wfn
#scf_e, scf_wfn = self.psi4.frequency(basis_sets, return_wfn=True, ref_gradient=self.wfn.gradient())
#print(scf_wfn.frequencies().get(0, 0))
except self.psi4.OptimizationConvergenceError as cError:
print('Convergence error caught: {0}'.format(cError))
self.wfn = cError.wfn
scf_energy = self.wfn.energy()
self.psi4.core.clean()
self.mol = self.xyz2mol()
if not self.debug:
self.psi4.core.opt_clean() # Seg fault will occured when the function is called before optimize.
return scf_energy
What Psi version are you using?
wying
January 5, 2023, 1:18pm
4
I am trying to calculate the binding energy of a cation and anion, which is the optimized energy of the pair minus of the optimized cation and anion. However, I either get the convergence error, or get positive binding energy, which is apparently not possible, since the binding energy of the cation anion pair should be negative for the attractive interaction between them. I need to getting back to the response of a paper under review. Thanks a lot!
I need the entire conda list - Psikit refuses to install on my Python 3.9 installation.
wying
January 6, 2023, 3:48am
6
_libgcc_mutex 0.1 main
wying
January 6, 2023, 3:49am
7
Actually, is this the verision issue? Can you calculate the energy for the sml structure I included based on your setup?
Thanks for your response!
The initial geometry from Psikit is very bad. There are two carbon atoms that are 0.879 A away from each other - completely unphysical.
I strongly recommend that you double-check your SMILES structure. If it still looks reasonable, move your issue to Psikit.
NOTE: A pre-edit version of this post implicated excited states. Upon doing another check, I don’t think that’s at issue here.
Even worse, there’s an N and a C that are less than 0.2 A apart. You need a better guess geometry.
wying
January 7, 2023, 4:38am
10
Thanks for your response. The psikit is using rdkit to directly transform the SMILE structure to geometry. In case of the individual cation and anion, the psi4 perform very well. I guess it is the distance between the cation and anion in the intital geometry is very bad. However, I currently cannot think about better way to optimize the geometry, since I am doing he high throughput calculation. Will increase the iteration number help the convergence?
With starting geometries that bad, maybe you could use steepest-descent with a Leonard-Jones potential - or something like tha? The Newton-Raphson-based molecular optimizers are not built to handle gracefully things like forces at 0.2 Angstroms.
wying
January 11, 2023, 8:43am
13
Seems like it doesn’t work, even through the iteration number set to 2000.
wying
January 11, 2023, 8:46am
14
Thanks for your comments. I am quite new to psi4, do you have any guidance to choose the Leonard-Jones potential in the optimization function?
This tool was developed for testing. I don’t know if it will be of practical use, but perhaps it at least gives you a start. At present the function (in optking/lj_functions.py) uses only one set of parameters.
In optking/tests/test_opthelper.py you will find a function test_lj_external_gradient() that shows how to explicitly loop over a L-J gradient from a function in optking.
If you are looking for an easy blackbox solution, you might look for a different software program instead built to generate initial guess structures.
wying
January 13, 2023, 2:53am
16
Thanks for your help! I will try Gaussian instead for now.