From: JC Gumbart (gumbart_at_ks.uiuc.edu)
Date: Mon Aug 20 2012 - 20:58:53 CDT
Yes, I'm trying to do these calculations over 1-2 ns trajectories, meaning I have only ~1000 frames at best. So there is no way the pressure calculation will be converged then.
Thanks for the help! I have enough to work with; even if the absolute numbers aren't reliable, I think the shapes of the profiles are.
JC
On Aug 20, 2012, at 11:22 AM, Grace Brannigan wrote:
> Hi JC,
>
> We post-processed the trajectories using the procedure you describe - one run for Ewald and one run for everything else, then added them together. We used the exact same configuration as for the run, (with the exception of the temperature) - and set the values of pressureProfileEwaldX,Y,Z to correspond to those of the PME in the simulation.
>
> We didn't get big fluctuations with T ~ 0, but we analyzed thousands of frames - I guess you have far fewer? We did find that with T~0 increasing the number of frames improved the situation far more than with T~300K. If the overall group pressure (i.e. the trace of the pressure tensor, output on-the-fly) isn't converged over the number of frames you have then yeah, I agree that it's unlikely the individual components will be converged, regardless of how you set the temperature in post-processing.
>
> cheers,
> Grace
>
> On Sat, Aug 18, 2012 at 7:24 PM, JC Gumbart <gumbart_at_ks.uiuc.edu> wrote:
> Sorry for taking so long to reply, but I've been running a battery of tests to try to get to the bottom of this.
>
> I had seen your previous post and was hoping you would chime in. The problem here is that I want the actual pressure profiles, not the surface tension (I was just trying that as a check, and I did try it as well with T = 0). So I can't just ignore the kinetic term.
>
> I ran my tests on a water box, calculating the profiles both during and post-processing. Through trial and error, I determined that the proper approach for post-processing is to run two calculations, both with the exact same configuration as the simulation itself (including both runs using PME), one to get all the contributions except for the Ewald part and one just to get Ewald. An important point I figured out here: the value of pressureProfileEwaldX,Y,Z really matters. When I used 20, it was much slower, but gave different results than 10. I tried 40, which is extremely slow, but is almost the same as 20, suggesting that using a value of 20 ensures this quantity is converged.
>
> I also ran a test for 100 steps, getting the pressures and the profiles with 1 slab every step. From this I was able to confirm what I found above. The sum of the concurrently run profile and the Ewald component with "pPEX,Y,Z 20" gives precisely the same value as the group pressure determined by NAMD. Post-processed (therefore with a different kinetic term) was similar, but with a small deviation as expected.
>
> Ultimately, what I'm concluding here is that the noise is primarily due to the fluctuations in the pressure calculations, not the kinetic term. Even at T=0 and a single slab, the fluctuations in the pressure profiles for the water box (and thus also the surface tension) are +/- 160 bars. So the profiles are not reliable, at least for so few samples, which is all I have from the DCDs.
>
> So Grace, when you did your calculations, did you 1) post-process the trajectories and 2) see similar fluctuations, even at T=0?
>
> Thanks!
> JC
>
> On Aug 16, 2012, at 2:33 PM, Grace Brannigan wrote:
>
>> Hi JC,
>>
>> I found that most of the noise in the pressure profiles was due to the ideal gas term, which is artificially noisy in post processing because velocities are randomly reassigned - see e.g.
>>
>> http://www.ks.uiuc.edu/Research/namd/mailing_list/namd-l.2009-2010/0250.html
>>
>> or
>>
>> http://www.ks.uiuc.edu/Research/namd/mailing_list/namd-l.2008-2009/4112.html
>>
>> My guess is that you will dramatically reduce noise if you post process at almost 0 degrees.
>>
>> cheers,
>> Grace
>>
>> On Mon, Aug 13, 2012 at 10:11 PM, JC Gumbart <gumbart_at_ks.uiuc.edu> wrote:
>> Does anyone have experience post-processing trajectories to get pressure profiles? I see here ( http://www.ks.uiuc.edu/Research/namd/2.9/ug/node75.html ) that I need to do separate calculations for the non-Ewald and Ewald contributions. One thing that is not obvious to me though is if I should turn off PME for the non-Ewald offline run. I would guess it should be left on, as one might typically calculate the non-Ewald component during the simulation itself, in which PME was on. Nonetheless, this assumption may or may not be valid.
>>
>> A bigger question though - given the large fluctuations in pressure, how useful is a completely offline calculation? Because I only saved my trajectory every 1000 timesteps, for a 1 ns run, I have only 1000 frames from which to calculate and average the profiles. So they are quite noisy. But more importantly, the surface tension does not approach 0 at the edges of my profile, but it does appear to level off to something (modulo the large fluctuations anyway). Maybe this is just due to insufficient sampling of the pressure, or maybe I'm doing something completely wrong?
>>
>> If anyone has done these calculations and gotten results that seem reasonable, please chime in with precisely how you did them.
>>
>> Thanks!
>> JC
>>
>>
>>
>> --
>> Grace Brannigan
>> Assistant Professor
>> Center for Computational and Integrative Biology (CCIB) &
>> Department of Physics
>> Rutgers University, Camden, NJ
>> (856)225-6780
>> www.branniganlab.org
>>
>
>
>
>
> --
> Grace Brannigan
> Assistant Professor
> Center for Computational and Integrative Biology (CCIB) &
> Department of Physics
> Rutgers University, Camden, NJ
> (856)225-6780
> www.branniganlab.org
>
This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:21:55 CST