Question about dtdmc

[Ivan_Veloso]

I am a new user in QMC and I use this program for a little time, for this reason I don't have much experience.

This last week I tried to obtain a dissociation energy with the casino program. First, I optimized the Jastrow parameters and then I calculated the DMC energy. To obtain the energy I used 100000 dmc steps and 2000 configurations for the statistics and I repeated this for different time step(0.001, 0.0025 and 0.005). The problem is when the time step is small(0.001) I represent the reblock graph for the dmc and I don't find the "plateau" at small values of dtdmc in the reblock.

Block len Std error Err in error
1 1.190778E-04 2.663150E-07
2 1.579759E-04 4.996579E-07
4 2.119098E-04 9.478781E-07
8 2.847128E-04 1.801075E-06
16 3.808577E-04 3.407375E-06
32 5.042606E-04 6.380601E-06
64 6.531861E-04 1.168938E-05
128 8.417852E-04 2.131036E-05
256 1.056928E-03 3.785678E-05
512 1.269994E-03 6.439546E-05
1024 1.481055E-03 1.064076E-04
2048 1.718183E-03 1.754235E-04
4096 1.815489E-03 2.638888E-04 *** BEST ***
8192 1.489433E-03 3.122125E-04
16384 1.728491E-03 5.359214E-04
32768 1.592540E-03 7.764297E-04

Then I try to increment the number of steps (600000) but I found a value very similar.

When I increase the value of dtdmc(0.0025 and 0.005) a "plateau" appear at high values(Blocklen=4096) even if I increase the number of steps to 800000. For this reason I want to know which value of dtdmc I have to use, because I don't know how to choose the right value. Should I increase more the number of steps to obtain a better reblock or I use the lowest value of dtdmc in which appear the "plateau"?.

[Ryan_Hunt]

Ivan,

Could you give more information about your specific system? Most importantly, are you using pseudo potentials or are you performing all electron calculations?

Thanks, Ryan.

[Ivan_Veloso]

Dear Ryan,

Now I work with two benzenes and I performing a all electron calculation. Do you need more information?

Ivan

[Ryan_Hunt]

Ivan,

I would say that your time steps seem reasonable then. Barring any malfunctions the only advice I can give is to accumulate more statistics. 800k steps for an AE dmc calculation is not massively big.

How many equilibration steps have you performed?

Are you running with an up-to-date version of CASINO? (I. E. Is it an old version that was compiled with gfortran?)

Thanks, Ryan.

[Ivan_Veloso]

Dear Ryan,

I use the version 2.12 of the CASINO program.

I used 5000 equilibration steps, because I read that is enough with a small number of steps. Anyway I will try to run the calculation with more statistics steps and I will check the results, but the problem is that I don't know how to choose the value for dtdmc.

Ivan.

[Ryan_Hunt]

Dear Ivan,

If you have compiled v2.12 of CASINO with gfortran, then be careful with your AE DMC calculations - we recently noticed a problem with the default gfortran compiler flags that leads to additional noise (and sometimes disastrous breaking of the variational principle) in AE DMC. This is fixed as of v2.13.599, but your version should still be bugged (check if the "-Ofast" flag appears in <CASINO BASE DIRECTORY>/arch/data/f90/gnu.arch - if it does, I would replace it with "-O3 -ffast-math -fstack-arrays" and recompile (and rerun your DMC...).

If you use a different compiler, then this wont be an issue .

With regards to choosing a sensible dtdmc and a reasonable equilibration period, I would recommend the following advice.

1) dtdmc should be chosen such that the root mean-squared distance diffused by an electron in one move (which is = sqrt(3 x dtdmc x p) with p = the step acceptance probability - usually modulated such that it is close to 1) is of order the smallest physical length scale in the problem (for AE calculations, this will be small, and hence dtdmc will typically be smaller than it would be if one used PPs for the same system).

2) the equilibration period should be chosen such that, on average, each electron in the system has chance to "properly explore the system" - if L is the longest length scale in your system, then you would want to have L ~ sqrt(3N_eq x dtdmc x p). Once you have decided on dtdmc via part (1), you can then find N_eq - hopefully a reasonable number of equilibration steps (something of order thousands is typical).

Hope this helps you!

Thanks, Ryan.