in support of my last suggestion, I plotted the x and y box size and
sqrt(x*y) for one of my simulations (one that completed, I should not have
overwritten the halted ones...) and it certainly seems the later term is
very constant while the x and y alone can vary more or less randomly by 10%
or more.
This would explain the seemingly random number of the crashes, as the box
dimensions change almost in a random walk. This also would mean that every
membrane simulation in pmemd.cuda should sooner or later crash with ntb=2
since nothing stops the x and y dimensions from sooner or later changing
too far.
Maybe I'm on the wrong track here, but if the above is correct, should
ntb=2 even be recommended for membrane simulations (which the manual does)?
Sooner or later, the box will thin so much that one will see boundary
artefacts.
Would ntb=3 then be preferred? But that requires surface tension > 0 and
lipid21 membranes should be run without surface tension afaik?
Kind Regards,
Thomas
On Mon, May 15, 2023 at 3:50 PM Thomas Steinbrecher <
thomas.steinbrecher.roche.com> wrote:
> Hi all,
>
> There is one more thing that occured to me: My system has ntb = 2 set,
> non-isotropic box scaling as recommended for membrane systems. Is it
> possible that the system "drifts", since the membrane can freely change
> shape in x-y direction as long as the total xy-area stays constant? This
> wouldnt even be wrong physically, but I can imagine pmemd could have an
> issue with that when splitting atoms into cells.
>
> There is no semi-isotropic scaling in Amber that varies the z-box size
> independent of the coupled x-y size, right?
>
>
>>
>> Just to check: how far into the simulation does this occur? Are you
>> printing often enough to be sure that you are seeing all the volume
>> changes that
>> are taking place? (Especially if the error is early in the run.)
>>
>
> it happens more or less equally spaced during my 250ns production run, as
> early as 30ns, or as late as 200ns into the run.
>
> I'm certainly not printing often enough to see quick changes, if something
> goes seriously wrong and destabilises the system in a few 1000s steps, I
> would likely miss it (Im printing every 1 ns, which already creates huge
> files).
>
>>
>> Note that vlimit is not checked on the GPU, but you can monitor the
>> temperature via mdout. A couple things to try:
>>
>> 1. Have you done enough initial NVT eqilibration before switching to
>> NPT?
>> (I don't know what "enough" is here, but try 10 ns or so.)
>>
>
> the NVT equilibration is fairly short, but at the point here, we have done
> 20+ns of various restrained and unrestrained NPT equilibration, plus
> typically more than 50ns of production runs. Density, Temperature and
> Pressure all look fine the entire time, so it would be surprising (and very
> worrying) if there could still be insufficient equilibration in the system
> somehow...
>
>
>> 2. Try NPT with the MC barostat to see if you get past this problem;
>> you
>> can switch to the Berendsen barostat later.
>>
>
> A past test had MC barostat and did not crash (N=1 observation), but I
> thought that is not recommended for membrane simulations currently?
>
>>
>> Of course, it would be good if pmemd reported more details about what is
>> going on -- at least exactly which dimension values were involved.
>>
>
> Yes, it would be helpful to know what pmemd expects here, or maybe to
> write out the final coordinates when this condition occurs.
>
>
>
>
> --
> *Dr. Thomas Steinbrecher*
> Principal Scientist CADD
>
> Roche Pharma Research and Early Development
> Roche Innovation Center Basel
> F. Hoffmann-La Roche Ltd
> Bldg. 092/3.92
> Grenzacherstrasse 124
> 4070 Basel
> Switzerland
>
> Phone +41 61 682 1319
> mailto: thomas.steinbrecher.roche.com
>
--
*Dr. Thomas Steinbrecher*
Principal Scientist CADD
Roche Pharma Research and Early Development
Roche Innovation Center Basel
F. Hoffmann-La Roche Ltd
Bldg. 092/3.92
Grenzacherstrasse 124
4070 Basel
Switzerland
Phone +41 61 682 1319
mailto: thomas.steinbrecher.roche.com
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Mon May 15 2023 - 07:30:02 PDT
