What is the best way of repeating simulations, say I have to run the 10 replicates of the same simulation trajectory which was previously simulated for 100 ns production run after minimization and equilibration steps. I took the last structure of the 100 ns simulation trajectory and used that last structure as the starting structure for the 10 replicates of the same system. The mdp file I am using for the 10 replicates has gen_vel=no but gen_seed=-1, will this parameters in the input file will produce new velocities? as I am seeing the trajectories are different from each other which would be good for statistical sampling and analysis etc also with the above mentioned setting I am not minimizing or equilibrating the replicates as I already took their starting structure from the well equilibrated system.
can anyone please tell me is this the correct way to produce replicates for the same system? As if have to do many replicates i.e more than 50 and each time I have to do gen_vel=yes with short equilibrations and before real production runs that will be time consuming and also will create a burden on server for job submission if I have to submit all 50 replicates together.
Dear Md Simulation
No, it does not generate new velocities. gen_seed=-1 means that the seed used is random, but if you are not generating velocities with gen_vel=yes then it is a useless specification.
About the sampling of the same systems to collect statistics, it depends also on the system itself. Sometimes it is fine to start from the same configuration and just produce different starting velocities by setting gen_vel=yes and gen_seed=-1. Another approach I have seen is to take a trajectory and extract frames at certain time intervals. For example, if you have a production trajectory of a system that has reached equilibrium, lets say 100 ns, you could extract one snapshot every 2ns and obtain a set of 50 replicas with different starting conditions.
Hope this helps,
Nicola
Nicola Piasentin Thank you for the reply , if every time having generated the velocities with gen_vel=yes parameter then one has to equilibrate each replica e.g 50 replicas 50 short equilibrations? and separate 50 production runs? because gen_vel = yes cannot be used for production runs, please correct me if I am wrong.
secondly I tested with gen_vel = no and gen_seed = -1 and created 30 replicas by taking one snapshot at 100th ns of the well equilibrated trajectory. I ran the simulations all of the trajectories have identical profiles (not exactly same their are differences among the profiles and trajectories) which I think is required for statistical sampling. Then at the end I average all of the replicates for the analysis purposes, this approach is not correct to make the replicates for having more data for statistical analysis etc? Please let me know.
Also I think if someone has to make replicates of the same trajectory it should not be created by generating new velocities as then it will be new trajectory because the velocities become regenerated then it will not be a replicate, replicates are require for increasing the statistical sampling instead of one trajectory one has 50 trajectories to perform the same analysis and average the results. So replicates must not be created with regenerating the velocities. Kindly correct me if I am thinking in a wrong direction.
I will be really grateful to you
Md Simulation
I would say that the point of not using gen_vel=yes for productions is that usually it sets the starting velocities, so it is implicit that you are basically at the first step of MD (energy minimization excluded). However, I would still run a short production to see if everything is okay. Honestly, I wouldn't start from the same point. You are kinda of relying on internal differences to get different samples, which can happen (https://www.researchgate.net/post/Why-does-performing-molecular-dynamics-simulation-on-different-CPUs-lead-to-different-results) but still I would prefer to run from different starting points rather than expecting them to diverge.
About the last comment you make, the velocities should not change the final results you get, that is, if you change the starting microstate (velocities) you should still end up at the same macrostate. Think about some general properties, for example the density of water. If you start from scratch by producing many boxes of water, and run MD on all of them at the same NpT conditions, then you would expect them to converge to the same macrostate where, for example, the average density is ca. 1 kg per cubic decimetre. Trajectories of water molecules from different systems will all be different, because all of these trajectories correspond to different points in the phase space given by the position and the velocities of the molecules, however all these microstates will correspond to the same macrostate.
Personally, I see the term replicate as 'same system/macrostate but different microstate' (although in some situations, line in umbrella sampling, some refers to replicas as the different windows, e.g. is same system with a molecule confined in a certain region of the reaction coordinate). In the case of the water boxes, all of them are the same system (# of water molecules, temperature and pressure if NpT, etc), but they will all start from different microstates (all positions and momenta of the molecules).
Think about your properties. You would like them to be general, right? And reproducible. If each of us, running a simulations in our PCs, would get a different result for the density, then we wouldn't be able to do any further step. Therefore, if I have many boxes with different starting velocities, I would still expect all of them to converge to the same averaged properties. The trajectories in the phase space (and therefore in GROMACS' .xtc file) will be different, but each system will (hopefully) end up in the correct ensemble.
Anyway, these are non trivial question, especially the reproducibility aspect of MD. I invite you to take my words as coming from another student in the field, without any claim of completeness or absolute truth. There are many books and papers on these topics, especially about statistical thermodynamics and MD, which should be hopefully more complete and clear many of your doubts!
Hope this helped a bit
Best
Nicola
- Badges
- Science topic