Solved problems on small systems following your advices, i am trying now the optimization of medium size crystals like quartz or alpha alumina (10 atoms primitive 3d cells, 2x2x2 correlation supercell).
Due to the small machine i use (48 processor), I have to take care about details, for getting some result within few days of cpu.
- are the jastrow parameters transferable?
So, for example:
Will I have some advantage in the vmc optimization time if i get jastrow-U parameters from an optimized vmc wave-funtion of another semi-ionic system?
Will I have some advantage if i get jastrow-Chi and jastrow-F parameters for Si in SiO2 from the optimized vmc wavefunction of Silicon?
thank you
strategies in vmc optimization
-
Neil Drummond
- Posts: 113
- Joined: Fri May 31, 2013 10:42 am
- Location: Lancaster
- Contact:
Re: strategies in vmc optimization
Dear Mauro,
Reusing the Jastrow u, chi and f terms should give a reasonable starting point. Note however that the cutoff length for u cannot exceed the Wigner-Seitz-cell radius. Hence if you are going from a larger simulation cell to a smaller cell you may need to reduce the cutoff length for u, in which case the Jastrow factor will no longer be good. Reusing a p term (or a u term in the presence of a p term) is not so straightforward, because the stars of G vectors are not necessarily the same when you go from one cell to another.
Best wishes,
Neil.
Reusing the Jastrow u, chi and f terms should give a reasonable starting point. Note however that the cutoff length for u cannot exceed the Wigner-Seitz-cell radius. Hence if you are going from a larger simulation cell to a smaller cell you may need to reduce the cutoff length for u, in which case the Jastrow factor will no longer be good. Reusing a p term (or a u term in the presence of a p term) is not so straightforward, because the stars of G vectors are not necessarily the same when you go from one cell to another.
Best wishes,
Neil.
-
Mauro_Causa
- Posts: 8
- Joined: Sun Jun 23, 2013 1:41 pm
Re: strategies in vmc optimization
Thank you Neil.
Following your suggestion I am trying to transfer jastrow parameters (u, chi, f)
They seems less transferable than another set of parameters i am used to, the gaussian basis sets, but it is another story
---------------
some question more:
The longer step in my vmc optimization is "derivative calculation":
Computing derivatives.
[CPU time: 2m37s elapsed, 9h remaining]
Done. [total CPU time: 8h45m]
- I didn't find much in the manual about: what are they? Derivative of the energy with respect the jastrow parameters?
- All the vmc optimization method use these derivative?
thank you
Following your suggestion I am trying to transfer jastrow parameters (u, chi, f)
They seems less transferable than another set of parameters i am used to, the gaussian basis sets, but it is another story
---------------
some question more:
The longer step in my vmc optimization is "derivative calculation":
Computing derivatives.
[CPU time: 2m37s elapsed, 9h remaining]
Done. [total CPU time: 8h45m]
- I didn't find much in the manual about: what are they? Derivative of the energy with respect the jastrow parameters?
- All the vmc optimization method use these derivative?
thank you
-
Neil Drummond
- Posts: 113
- Joined: Fri May 31, 2013 10:42 am
- Location: Lancaster
- Contact:
Re: strategies in vmc optimization
Dear Mauro,
It is evaluating numerical derivatives of the local energies with respect to wave-function parameters.
If you are just optimising a Jastrow factor and you have fixed the cutoff lengths, you can use opt_method=varmin_linjas. This is an optimised variance minimisation algorithm for the linear parameters in the Jastrow factor, which is much faster than the other optimisation methods. It does not require numerical derivatives.
Best wishes,
Neil.
It is evaluating numerical derivatives of the local energies with respect to wave-function parameters.
If you are just optimising a Jastrow factor and you have fixed the cutoff lengths, you can use opt_method=varmin_linjas. This is an optimised variance minimisation algorithm for the linear parameters in the Jastrow factor, which is much faster than the other optimisation methods. It does not require numerical derivatives.
Best wishes,
Neil.