From: Axel Kohlmeyer (akohlmey_at_gmail.com)
Date: Sat May 16 2015 - 00:06:27 CDT
On May 16, 2015 4:08 AM, "Grzegorz Nawrocki" <aksonik_at_gmail.com> wrote:
>
> Dear Bradley and Axel,
>
> First of all, thank you very much for your quick responce. For me it is
just the beginning of using NAMD. I am grateful for your help. Second, I
heard about "butterfly effect" and I know that the distribution of the
various states is the usable output. However, I expected that no
differences should exist, even tiny, for exact restart. I am also aware of
introducing "randomness" into the system by using temperature control and
setting seed, but I do it only at the beginning, and the beginning is the
same for the simulations with and without restarting. Providing coordinates
and velocities I should be able to get exactly the same trajectory from any
checkpoin. Where do these tiny differences come from?
Like I already told you, they do come from using floating point math and
the langevin thermostat.
>
> Best regards,
> Grzegorz
>
> 2015-05-15 19:26 GMT-04:00 Axel Kohlmeyer <akohlmey_at_gmail.com>:
>
>> On Fri, May 15, 2015 at 7:09 PM, Grzegorz Nawrocki <aksonik_at_gmail.com>
wrote:
>> >
>> > Dear NAMD users,
>> > I have noticed that NAMD does not generate the same trajectory when a
>> > simulation is run with and without restarting. I tested it using very
simple
>> > system - it consists of two atoms. I used a single core, temperature
>> > coupling, constant volume and mostly default parameters. Please find
the
>> > enclosed input files for a short simulation.
>> >
>> > Without restarting NAMD everytime generates the same trajectory, as
well as
>> > with restarting, but both trajectories, i.e. with and without
restarting,
>> > are becoming more and more different with every restart.
>> >
>> > Is it a bug or I set something wrong?
>>
>> MD is chaotic and thus trajectories *have* to exponentially diverge
>> over time when there is the tiniest difference (ever heard the term
>> "butterfly effect"?). restarting enforces such tiny differences. the
>> usable output of MD simulations are not the exact positions anyway,
>> but the *distribution* of the various states that are visited over
>> time. you can look up details in text books on statistical mechanics
>> and thermodynamics.
>>
>> axel.
>>
>> p.s.: you *can* have an MD code that produces exactly reproducible and
>> reversible trajectories, but then you'll have to do: 1. fixed point
>> math, 2. only time integration and not thermostat, 3. fixed volume.
>>
>>
>>
>> >
>> > Best regards,
>> > --
>> > Grzegorz Nawrocki
>>
>>
>>
>> --
>> Dr. Axel Kohlmeyer akohlmey_at_gmail.com http://goo.gl/1wk0
>> College of Science & Technology, Temple University, Philadelphia PA, USA
>> International Centre for Theoretical Physics, Trieste. Italy.
>
>
>
>
> --
> Grzegorz Nawrocki
This archive was generated by hypermail 2.1.6 : Thu Dec 31 2015 - 23:21:52 CST