Optimization problem of the AE CUTOFFS

 Posts: 7
 Joined: Tue Mar 04, 2014 11:07 am
Optimization problem of the AE CUTOFFS
I am facing an optimization problem of the AE CUTOFFS (They are not being optmized) using the CASINO version 2.13.584. I have been previously working with the version 2.12.1 which works properly. Can anyone help me with this?

 Posts: 52
 Joined: Thu Jan 30, 2014 1:25 am
Re: Optimization problem of the AE CUTOFFS
Hi Braulio,
If I understand correctly you are trying to optimize the backflow AE cutoffs but their values do not change. Could you post your input / correlation.data / out files to see if there are any obvious problems?
Best,
Pablo
If I understand correctly you are trying to optimize the backflow AE cutoffs but their values do not change. Could you post your input / correlation.data / out files to see if there are any obvious problems?
Best,
Pablo
Hey there! I am using CASINO.

 Posts: 7
 Joined: Tue Mar 04, 2014 11:07 am
Re: Optimization problem of the AE CUTOFFS
Hi Pablo,
The files asked are attached.
Best regards,
Braulio Gabriel.
The files asked are attached.
Best regards,
Braulio Gabriel.
 Attachments

 ProblemCasino.tar.gz
 (246.22 KiB) Downloaded 858 times

 Posts: 52
 Joined: Thu Jan 30, 2014 1:25 am
Re: Optimization problem of the AE CUTOFFS
Hi Braulio,
For some reason the AE cutoffs are being initialized to 1.1E8 a.u., which is essentially impossible to optimize. Add the following to your initial correlation.data just before the "END BACKFLOW" label:
which will not only set the cutoffs to reasonable initial values but will also use the same L_AE for the three nuclei in your system.
I will look into why the code came up with such poor initial guesses for the cutoffs.
Hope this helps,
Pablo
For some reason the AE cutoffs are being initialized to 1.1E8 a.u., which is essentially impossible to optimize. Add the following to your initial correlation.data just before the "END BACKFLOW" label:
Code: Select all
START AE CUTOFFS
Nucleus ; Set ; Cutoff length ; Optimizable (0=NO; 1=YES)
1 1 1.d0 1
2 1 1.d0 1
3 1 1.d0 1
END AE CUTOFFS
I will look into why the code came up with such poor initial guesses for the cutoffs.
Hope this helps,
Pablo
Hey there! I am using CASINO.

 Posts: 7
 Joined: Tue Mar 04, 2014 11:07 am
Re: Optimization problem of the AE CUTOFFS
Thank you Pablo. I'll do some testing.

 Posts: 7
 Joined: Tue Mar 04, 2014 11:07 am
Re: Optimization problem of the AE CUTOFFS
I found something different in the optimized coefficients of the AE CUTOFFS. They have the same values.
START AE CUTOFFS
Nucleus ; Set ; Cutoff length ; Optimizable (0=NO; 1=YES)
1 1 0.878846061910215 1
2 1 0.878846061910215 1
3 1 0.878846061910215 1
END AE CUTOFFS
START AE CUTOFFS
Nucleus ; Set ; Cutoff length ; Optimizable (0=NO; 1=YES)
1 1 0.878846061910215 1
2 1 0.878846061910215 1
3 1 0.878846061910215 1
END AE CUTOFFS

 Posts: 52
 Joined: Thu Jan 30, 2014 1:25 am
Re: Optimization problem of the AE CUTOFFS
Hi Braulio,
Yes, I mentioned this in my previous post  the label "1" in the specification of the AE CUTOFFS forces them to be the same, so there are two fewer parameters to worry about. In your correlation.data file you force the parameters in the MU and PHI backflow terms to be the same for the three atoms in your system (by the specification of "Atoms in this set: 1 2 3"), so it makes sense for the AE CUTOFFS to be forced to be equal too.
In any case, if this is not what you want you can just set a different label for each cutoff:
Best,
Pablo
Yes, I mentioned this in my previous post  the label "1" in the specification of the AE CUTOFFS forces them to be the same, so there are two fewer parameters to worry about. In your correlation.data file you force the parameters in the MU and PHI backflow terms to be the same for the three atoms in your system (by the specification of "Atoms in this set: 1 2 3"), so it makes sense for the AE CUTOFFS to be forced to be equal too.
In any case, if this is not what you want you can just set a different label for each cutoff:
Code: Select all
START AE CUTOFFS
Nucleus ; Set ; Cutoff length ; Optimizable (0=NO; 1=YES)
1 1 1.d0 1
2 2 1.d0 1
3 3 1.d0 1
END AE CUTOFFS
Pablo
Hey there! I am using CASINO.

 Posts: 7
 Joined: Tue Mar 04, 2014 11:07 am
Re: Optimization problem of the AE CUTOFFS
Thanks Pablo for the explanation.

 Posts: 138
 Joined: Wed Apr 15, 2015 3:14 pm
Re: Optimization problem of the AE CUTOFFS
Hello all.
I noticed that when one optimize backflow parameters using VMC emin, the AE CUTOFFS parameters is bad or not optimized at all i.e. accepts random values or not optimal in terms of subsequent DMC calculation.
As stated in APPENDIX B: ZEROING THE BACKFLOW DISPLACEMENT AT AE ATOMS in article Inhomogeneous backflow transformations in quantum Monte Carlo P. Lopez Rıos, A. Ma, N. D. Drummond, M. D. Towler, and R. J. Needs
With large values of the parameter the energy becomes worse in the subsequent DMC calculation, and with smaller values, the variance in VMC emin optimization increases and it is difficult to obtain reproducible values for optimal backflow parameters.
Although if one increase the number VMC configurations up to 10^610^7 with decreasing AE CUTOFFS to 0.01 and optimize subsequently one can get a better DMC energy value. This is true for both atoms and molecules and DMC energy get better roughly 1mHa per atom when AE cutoff decrease from 0.5 to 0.01 (for hydrogen atoms did not check).
I have a question how effective the formula (B1 from APPENDIX B) is for small AE CUTOFFS (order of 0.01 au) and whether it can be improved without complicating the calculations?
Best Vladimir.
PS.
I found http://www.lancaster.ac.uk/staff/drummo ... apers.html is very usefull for citing
PPS.
I use ETA term of order = 9
I noticed that when one optimize backflow parameters using VMC emin, the AE CUTOFFS parameters is bad or not optimized at all i.e. accepts random values or not optimal in terms of subsequent DMC calculation.
As stated in APPENDIX B: ZEROING THE BACKFLOW DISPLACEMENT AT AE ATOMS in article Inhomogeneous backflow transformations in quantum Monte Carlo P. Lopez Rıos, A. Ma, N. D. Drummond, M. D. Towler, and R. J. Needs
Manually selecting the values of the AE CUTOFFS parameters, I found that the optimal values fall into a range of 0.20.5 au.AE cutoff limit the radii of nonintersecting spheres around the atoms inside which an artificial multiplicative cutoff function is applied to all contributions to the backflow displacement of particle that do not depend on the distance to the central AE atom. This includes the homogeneous backflow displacement and the inhomogeneous contributions from neighboring atoms.
With large values of the parameter the energy becomes worse in the subsequent DMC calculation, and with smaller values, the variance in VMC emin optimization increases and it is difficult to obtain reproducible values for optimal backflow parameters.
Although if one increase the number VMC configurations up to 10^610^7 with decreasing AE CUTOFFS to 0.01 and optimize subsequently one can get a better DMC energy value. This is true for both atoms and molecules and DMC energy get better roughly 1mHa per atom when AE cutoff decrease from 0.5 to 0.01 (for hydrogen atoms did not check).
I have a question how effective the formula (B1 from APPENDIX B) is for small AE CUTOFFS (order of 0.01 au) and whether it can be improved without complicating the calculations?
Best Vladimir.
PS.
I found http://www.lancaster.ac.uk/staff/drummo ... apers.html is very usefull for citing
PPS.
I use ETA term of order = 9
In Soviet Russia Casino plays you.