Autocorrelation in gmx analyze

Hi All,

Simply I want to know whether showing the reaching of the autocorrelation function of the radius of gyration to zero using the last 200 ns of the simulation time (I simulated my system for 700 ns) can be used to imply the reaching of the equilibration of my system.

OR

When calculating the autocorrelation function of the radius of gyration to show the reaching of the system’s equilibration, do I have to use always data from t=0 to t=700 ns in my system?

Appreciate your help in this regard.

Thank you.

Hi,

I’m not sure exactly what you mean. If you think that it takes 500 ns (of the simulation time) to reach full equilibration, then I suggest only using the last 200 ns for analysis. But I would then also run a separate analysis on the whole 700 ns to see if there is any change over that period of time, especially during the last half of the simulation.

You certainly do not have to use data starting from t=0 if you believe that the initial part of the simulation is to be considered equilibration.

Thank you very much for your answer. Now I got it.

@MagnusL

Hello again,

I have a quick question. When calculating ACF, half of the frames are used by default for the length of ACF. For example, for a 700 ns simulation, the first 350 ns is used as the length of ACF. Can I use all frames for calculating ACF (why the default is set as half of the frames)?

On the other hand, if I use half of the frames (0-350 ns) for ACF calculation and if ACF reaches 0 around 300 ns (reaching equilibration), does this imply that 300-700 ns can be used for analysis?
On the other hand, if I use half of the frames (0-350 ns) for ACF calculation and if ACF reaches 0 around 300 ns (reaching equilibration), does this imply that 300-700 ns can be used for analysis?

You use all frames (from -b to -e) to calculate the ACF. The length of the ACF will be half that time. You can use a shorter length for the autocorrelation function, but I don’t think using a longer makes sense.

I am not sure you can use the time for your ACF to reach 0 as a basis for when equilibrium is reached. Please provide a reference.

@MagnusL

Thank you very much for your reply and clearing my doubts about ACF length. By referring to several papers, now I know that reaching ACF to zero implies having independent samples after 300 ns. Previously, I was confused with one of the replies in Autocorrelation function of radius of gyration - #5 by MOH4

By calculating the block average with error estimation (gmx analyze with -ee option), can we show the convergence of an MD system (observing a plateau in error estimate vs block size graph) ? For the calculation of the block average, do we need to have independent samples? Appreciate it if you could comment on this.

To get good statistics of independent samples you should have the samples spaced by 2 tau (with tau being the autocorrelation time), see Simulation Information Gathering - AlchemistryWiki, for a good summary.

Since you have a very long autocorrelation time for the RoG, I would use some other observable to see if your system is equilibrated. I’d recommend checking, e.g., that both the potential energy and the system volume are (relatively) constant in your production simulation, or the part of it that you use for measuring.

Thanks a lot for your answer.