[gmx-users] Different results from identical tpr after MD

[chris.neale at utoronto.ca]

Dear Acoot:

The idea of convergence is this: start a large number of simulations  
from different conformations, analyze some quantity over time in each  
simulation, and when the deviation of the average value of that  
quantity from each separate simulation is less than the time-variance  
within individual simulations, then you can imply that the simulations  
have converged -- that is, the results of independent simulations  
which started as different are now similar.

There is a huge body of work that uses a single simulations and  
evaluates its so-called convergence using some assumptions and special  
methods. That can also be very useful, but I find it informative to  
think of "convergence" in its standard non-scientific dictionary  
definition as the coming together of previously disparate things.

For simulations, my working definition is this: a set of simulations  
has converged the value of some variable when the simulations were  
initiated from sufficiently distinct conformational basins and then,  
over time, the ensemble distribution of the time-averages of the  
specified variable has a variance that is the same as the mean  
time-averaged variation within independent simulations. The weak point  
here is the part about "sufficiently distinct conformations", but I am  
not sure that this can be stated less vaguely in the general case.

Chris.

-- original message --

Hi Justin,

Can you give me your definition of converged MD and unconverged MD?

Cheers,

Acoot