Bias temperature issue in WT-metadynamics

900 views
Skip to first unread message

vikas....@gmail.com

unread,
Dec 4, 2016, 12:57:34 PM12/4/16
to PLUMED users
Hi everyone, 

 I use NAMD to run WT-MetaD simulations. I did WT- simulations at 3 different bias temperatures ( 180K, 300K, 1000K). Normal temperature is 310 K, which makes bias factor ( 1.58, 2.033, 4.22). I get different results everytime. I agree high bias temperature will help me sample cross longer hills but I am able to sample whole region with each of these bias temperatures, which  really makes me confused which result to trust. Could anyone give an answer how to go with selecting bias temperature ? and what are problems with choosing high bias temperature or may be direct me to some publication where affect of bias temperature is studied ? I will be extremely thankful.    



Thanks,
Vikas

Giovanni Bussi

unread,
Dec 5, 2016, 4:34:45 AM12/5/16
to plumed...@googlegroups.com
Hi,

the bias factor parameter allows you to tune the region of your free energy landscape you will explore.

In principles, converged simulations should give the same result independently of its choice. Thus the only explanation is that your simulations are not converged. Notice that:
1. Too low bias factor might not enhance sampling enough. In the limit of factor = 1 you have no metadynamics bias at all.
2. Too high bias factor makes the region to explore very large, and thus potentially more difficult to converge.

From your description it looks like all the simulations are exploring the same region. I guess that even if you run a new simulation at the same bias factor that one of those you run already you would get yet another result. In case this is true, that's just a confirm that metadynamics is not converging.

If so, the best you can get out of your MD is a qualitative insight of which are the stable conformations accessible and perhaps you can try to spot out which other observables are changing irreversibly during the simulation so as to perhaps add them as additional CVs.

Giovanni


--
You received this message because you are subscribed to the Google Groups "PLUMED users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to plumed-users+unsubscribe@googlegroups.com.
To post to this group, send email to plumed...@googlegroups.com.
Visit this group at https://groups.google.com/group/plumed-users.
To view this discussion on the web visit https://groups.google.com/d/msgid/plumed-users/fa60fc9f-3ac9-40f3-b87f-aa408897b572%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Vikas Dubey

unread,
Dec 5, 2016, 5:02:48 AM12/5/16
to plumed...@googlegroups.com
Hi, 

Thanks for your answer ! I am doubtful about the definition of convergence. From my understanding hill-height should go to zero. To check the convergence I am plotting collective variable w.r.t hill height. 
Here is the convergence plot at 180K (bias factor = 1.58). I felt its more or less converged. How relevant is plotting no of steps with hill height as mentioned in tutorial in case of 2-CV ? I feel  this gives me better idea of convergence.  
Inline image 2 


Thanks,
Vikas





--
You received this message because you are subscribed to a topic in the Google Groups "PLUMED users" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/plumed-users/pSibQNI0WP8/unsubscribe.
To unsubscribe from this group and all its topics, send an email to plumed-users+unsubscribe@googlegroups.com.

To post to this group, send email to plumed...@googlegroups.com.
Visit this group at https://groups.google.com/group/plumed-users.

Giovanni Bussi

unread,
Dec 5, 2016, 5:16:35 AM12/5/16
to plumed...@googlegroups.com
Hi,

as it has been written several times in this list, checking that hill's height goes to zero is not a good check for convergence.

A qualitative way is to check if there are multiple recrossings even in the last part of the simulation. More quantitatively you should check that the same regions in the CV space are visited at different stages of the simulation. There is an example in the review a wrote with Davide Branduardi on Rev Comput Chem, but I am sure many other reviews touch this point.

What I am recently doing to have a quantitative handle on convergence is the following:
1. Compute free energy using histograms with last bias (as explained e.g. in Branduardi, Bussi, Parrinello, JCTC 2012). Alejandro Gil-Ley wrote a brief explanation about how to use plumed to do it here: https://groups.google.com/d/msgid/plumed-users/d1219a72-e80b-4a45-af10-56d6bc15779a%40googlegroups.com?utm_medium=email&utm_source=footer 
You can also use a time averaged bias instead of the final one. Time averaging can be done as in Micheletti, Laio, Parrinello, PRL (2005).

2. Compute the error on the histogram with block analysis, that is: do histogram with one block, then redo it with another block, etc, then take error of the mean. You can also discard the first part of the simulation in accumulating the histogram. Notice that the first part of the simulation is the one where typically you see the larger number of spurious recrossing, just induced by over-destabilization of one state.

In my experience, whenever you have pathological cases the error comes out very large and act as a quantitative warning.

Notice that any other reweighting method could be used for step 1. My guess is that the performance should be similar.

Giovanni


Pratyush Tiwary

unread,
Dec 5, 2016, 11:11:56 AM12/5/16
to plumed...@googlegroups.com, Vikas Dubey
Dear Vikas

Just to complement what Giovanni already said, an unbiased MD simulation is essentially a metadynamics with  hill height =0 /  bias factor = 1. Thus the hill height is perfectly converged per construction. 

Does this mean the sampling is converged? Of course not!! 

Metadynamics has a variety of tools to construct the histogram as a function of simulation time. There are countless reviews out there - see for example www.annualreviews.org/doi/abs/10.1146/annurev-physchem-040215-112229

One necessary but not sufficient condition to check convergence is to get the free energy using two different methods - for example directly from the hills (F=-V type) or using reweighting (described in above review, implemented in PLUMED but otherwise very easy to code too for NAMD-colvars). If they agree, you might be on the right path. If they don't agree, your simulation needs to run longer or you need to think more about which collective variable to bias.

It might be a good idea to check out metadynamics for alanine dipeptide with various biasing parameters and familiarize yourself with the reweighting methods. The PLUMED tutorials provide information on this. You could patch your NAMD with PLUMED to directly use these tutorials, but even directly with the colvars module in NAMD which I think you are using, the knowledge of the tutorials is quite transferable.

Regards
Pratyush


--
Pratyush Tiwary
Post-doctoral Scientist, Columbia University
Department of Chemistry

On Mon, Dec 5, 2016 at 5:15 AM, Giovanni Bussi <giovann...@gmail.com> wrote:
Hi,

as it has been written several times in this list, checking that hill's height goes to zero is not a good check for convergence.

A qualitative way is to check if there are multiple recrossings even in the last part of the simulation. More quantitatively you should check that the same regions in the CV space are visited at different stages of the simulation. There is an example in the review a wrote with Davide Branduardi on Rev Comput Chem, but I am sure many other reviews touch this point.

What I am recently doing to have a quantitative handle on convergence is the following:
1. Compute free energy using histograms with last bias (as explained e.g. in Branduardi, Bussi, Parrinello, JCTC 2012). Alejandro Gil-Ley wrote a brief explanation about how to use plumed to do it here: https://groups.google.com/d/msgid/plumed-users/d1219a72-e80b-4a45-af10-56d6bc15779a%40googlegroups.com?utm_medium=email&utm_source=footer 
You can also use a time averaged bias instead of the final one. Time averaging can be done as in Micheletti, Laio, Parrinello, PRL (2005).

2. Compute the error on the histogram with block analysis, that is: do histogram with one block, then redo it with another block, etc, then take error of the mean. You can also discard the first part of the simulation in accumulating the histogram. Notice that the first part of the simulation is the one where typically you see the larger number of spurious recrossing, just induced by over-destabilization of one state.

In my experience, whenever you have pathological cases the error comes out very large and act as a quantitative warning.

Notice that any other reweighting method could be used for step 1. My guess is that the performance should be similar.

Giovanni

On Mon, Dec 5, 2016 at 11:02 AM, Vikas Dubey <vikas....@gmail.com> wrote:
Hi, 

Thanks for your answer ! I am doubtful about the definition of convergence. From my understanding hill-height should go to zero. To check the convergence I am plotting collective variable w.r.t hill height. 
Here is the convergence plot at 180K (bias factor = 1.58). I felt its more or less converged. How relevant is plotting no of steps with hill height as mentioned in tutorial in case of 2-CV ? I feel  this gives me better idea of convergence.  
Inline image 2 


Thanks,
Vikas




Reply all
Reply to author
Forward
0 new messages