(In any case let me try your factor idea, some other stuff that came to mind + finish reading your paper so that maybe I have more useful info on the problem) Em quarta-feira, 26 de abril de 2023 às 14:05:06 UTC+2, Cecília Álvares escreveu:
> Indeed, this could be the reason why I have this weird non-smoothness in > the plots I sent in my 2nd message (the ones concerning a less coarsened > mapping), because indeed in this case I was optimizing all the three bonded > potentials at once. I will try not doing them at the same time and see if > the smoothness-issue improves. > > But then this would not explain the issues I had in the original post I > made, which concerned another mapping (a highly coarsened one). If the > problem was a matter of optimizing more than one bonded potential at once, > I should have had good results when I tried to do IBI only for one angle > type and kept the potential for bonds constant (at a BI guess) throughout > the procedure. But unfortunately my angle distribution still converges to > something ultra weird with 3 peaks. > > PS: maybe my last message was too big and maybe it was confusing, but the > figures I sent in my 1st message and in my 2nd message are for different > mappings. In the first one (let's call it mapping A), I have only 1 bond > type and 1 angle type. For this one I did try optimizing separately to see > if it would fix the problem and yet I reached weird results. The second > message had figures of a less coarsened mapping (let's call it mapping B) > in which I somewhat successfully converge to potentials that yield more or > less rightful distributions (apart from the smoothness issue). I only > brought up the results of the second mapping to show that the same strategy > "worked" for deriving bonded potentials via IBI for another mapping. Sorry > if I made it more confusing! > > Em quarta-feira, 26 de abril de 2023 às 08:29:58 UTC+2, Marvin Bernhardt > escreveu: > >> Hey Cecília, >> >> Oh ok, then it is probably not the interaction with the non-bonded terms, >> that causes issues. But I believe something similar is going on, that >> indeed has something to do with your system being a solid/crystal: >> IBI is a very good potential update scheme, when the degrees of freedom >> are well separated. For molecules in liquids, angles and bonds are usually >> well separated, i.e. changing the potential of one, does not affect the >> dist of the other much. But multiple occurrences of equivalent DoFs also >> need to be well separated for IBI to work well. In your case, consider a >> single angle potential between three beads in the crystal is changed, but >> all the others are kept constant. It will change the distribution of that >> angle, but also have effect on different angles. In that case IBI is not >> providing a good potential update at each iteration. >> What is happening in detail, I believe, is that the angle potential of >> all angles is updated by IBI, but this leads to an “overshoot”. The next >> iteration, IBI tries to compensate, but overshoots again in the other >> direction. You can easily test if this is what is happening, plotting even >> and uneven iterations separately, i.e. compare a plot at iterations 10, 12, >> 14 with 11, 13, 15. >> This has happened to me before with ring molecules, where the situation >> is similar. A possible solution is to scale the update, by some factor >> between 0 and 1 (I'd try 0.25). >> >> Also test this for the bond potential, I guess this is happening there >> too, otherwise it should converge within ~20 iterations. >> >> Greetings, >> Marvin >> On Tuesday, 25 April 2023 at 10:25:59 UTC+2 Cecília Álvares wrote: >> >>> Hey Marvin, >>> >>> Thanks a lot for the reply! >>> I will have a look on the paper right now and do some thinking. In fact, >>> I wanted to test the possibility of optimizing the bonded potentials first >>> and, after its optimization is done, optimize the non-bonded. So basically >>> there is no optimization of non-bonded whatsover being done in my >>> simulation. To build the target distributions, I sampled an atomistic >>> system in which the non-bonded forces were artificially removed. After >>> having a trajectory file of this AA system, I built the corresponding >>> target distributions to be used in VOTCA with csg_stat. For what is worth >>> it, the target distributions of angle and bond don't seem at all weird >>> relative to the "real ones", of when non-bonded forces exist. And then, >>> after having the target distributions, I set up the CG MD simulations >>> within the IBI to have only bonded potential also. So, besides there being >>> no non-bonded potential optimization, there is also no non-bonded forces at >>> all in my CG system. But I dont think this should be a problem, right? It >>> makes sense to entrust the CG bonded potentials to reproduce the target >>> distributions of the AA bonded potentials. >>> >>> What I did try also, and that is in allignment with your idea, was to >>> set up two IBI runs: (1) one run to optimize *only* the potential for >>> the bonds and keep the angle potential active (in this case the latter >>> comes from a simple BI) and (2) one run to optimize only the potential for >>> the angles and keep the bond potential active (in this case the latter >>> comes from a simple BI). In the case (1) it seems that I converge to a >>> potential for bonds that is quite able to reproduce the corresponding >>> distribution, while in the case (2) I converge more and more to potentials >>> that give super weird distributions (like with three weird peaks, as I >>> showed in the figure above) >>> >>> Concerning the phase of the system: it is a solid system. More >>> specifically, it is a coarsened grained version of ZIF8 in which the whole >>> repeating unit was assumed to be one bead. I know that IBI has not at all >>> been developed for solids and even further not for MOFs - the goal is >>> actually to derive potentials in the CG level using many different >>> strategies (IBI, FM, relative entropy) and evaluate the results. In any >>> case, I dont think that the fact that my system is a xtalline solid could >>> be the reason why my results are super weird (right?). It seems like such a >>> simple system when in the CG level. >>> >>> For what is worth it, I am also assessing different mappings. Following >>> the same strategy of optimizing first bonded-potential for a less coarsened >>> mapping (2 beads), I am able to reach less weird results. For example, you >>> can find below the evolution of the corresponding distributions as I >>> perform more iterations for this system (it has one bond type and two angle >>> types). I think there is still a problem since we can see some tendency of >>> the distributions becoming non-smooth as I do more iterations, but the >>> results are definitely less weird. >>> >>> [image: picture.png] >>> >>> Em segunda-feira, 24 de abril de 2023 às 20:50:14 UTC+2, Marvin >>> Bernhardt escreveu: >>> >>>> Hi Cecília, >>>> >>>> I once encountered similar problems with bonded and non-bonded >>>> interactions. See Fig. 9 of this paper >>>> <https://pubs.acs.org/doi/10.1021/acs.jctc.2c00665>. In short: The >>>> problem was that the potential update of the non-bonded has some influence >>>> on the bonded distribution, and vice versa. But the potential update is >>>> calculated as if they were independent. >>>> >>>> The fix in my case was to update the two interactions alternately using >>>> `<do_potential>1 0</do_potential>` for bonded and `<do_potential>0 >>>> 1</do_potential>` for non-bonded interactions. You could try something >>>> similar. >>>> >>>> Otherwise, is your system liquid? Are there non-bonded interactions >>>> that you are optimizing at the same time? >>>> >>>> Greetings, >>>> Marvin >>>> >>>> On Monday, 24 April 2023 at 16:56:42 UTC+2 Cecília Álvares wrote: >>>> >>>>> Hey there, >>>>> >>>>> I am currently trying to derive bonded potentials of a very simple CG >>>>> system (containing only one bond type and one angle type) using IBI. >>>>> However, I have been failing miserably at doing it: instead of reaching >>>>> potentials that are better and better at reproducing the target >>>>> distributions for the bond and for the angle, I end up having weider and >>>>> weider distributions as I do the iterations. I am posting a plot of the >>>>> bond and angle distributions to give a glimpse on the "weirdness". I have >>>>> already tried: >>>>> (1) providing very refined (small bin size and a lot of bins) target >>>>> distributions of excelent quality (meaning not noisy at all) for the bond >>>>> and the angle. Similarly, I have also tried using less refined target >>>>> distributions (larger bin sizes and less amount of bins). >>>>> (2) varied a lot the setup in the settings.xml concerning bin sizes >>>>> for the distributions to be built at each iteration from the trajectory >>>>> file. I have tried very small bin sizes as well as large bin sizes. >>>>> (3) increasing the size of my simulation box hoping that maybe it was >>>>> all a problem of not having "enough statistics" to build good >>>>> distributions >>>>> at each iteration within the trajectory file I was collecting from my >>>>> simulations. >>>>> >>>>> None of these things has worked and I think I ran out of ideas of what >>>>> could possibly be the cause of the problem. Does anyone have any insights? >>>>> >>>>> I am also attaching my target distributions (this is the scenario in >>>>> which I am feeding target distributions lot of points and smaller bin >>>>> size) >>>>> and the settings.xml file for what is worth it. >>>>> >>>>> [image: plots.png] >>>>> >>>> -- Join us on Slack: https://join.slack.com/t/votca/signup --- You received this message because you are subscribed to the Google Groups "votca" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/votca/a6fd2f9e-3ae5-4fb2-919b-03d0e9dc674cn%40googlegroups.com.
