I setup two independent runs using AWH with pull code. I simply ran simulations independently, not using -multidir. The two runs have the same setup. I got two sets of energy files, and pullx.xvg files. Is it possible analyze these two sets of tiles together by combining the energy files using gmx eneconv?
Technically it is doable. But since the two energy files are not sequential in time, I want to make sure combining this way would not distort analysis.
If they are independent it’s better to use that as a feature rather than trying to combine them, i.e. so that you can find the variance/std dev./std. err. of the results. Preferably, you’d want at least four independent runs to get reasonable statistics.
I don’t see what you want to do with the energy files. Samples that you take for any observable over time do not have the same weight, as you have biased the simulation. So you can not directly take ensemble averages. Also you should ignore any part of the trajectory where the free-energy has not yet converged.
Hi Hess, thanks for the reply. Initially I wanted to do something similar as mdrun -multidir to collect multiple samplings in parallel. The HPC I am using does not have external MPI libraries installed, so I cannot use mdrun -multidir.
Also just naively wonder that if it is possible to let AWH sample specific segments of a reaction coordinate, then combine trajectories of these samples, in a way similar to umbrella sampling, then reconstruct the PMF? For umbrella sampling, I have trouble picking appropriate restraining force parameters. Sorry if the question sounds too naive. Thanks in advance!
Yes, you can actually speed up the sampling by cutting you interval into parts, as that avoids that individual replicas have to cover the whole interval, which in the end is a diffusive process. But that only works well when you have a single transition pathway (just as with umbrella sampling). Make sure that you have some overlap between the sub-intervals, such that you can reliably stitch together the PMFs.