Molecular Dynamics using machine learning is reproducible for same initial files.

[mitanshi_gupta]

I am training a machine-learned force field (MLFF) molecular dynamics simulation in VASP at 750 K with NSW=20000 steps for the training run, followed by a production run of NSW=20000 steps at different temperatures. I want to stabilize the structure, but I am getting different results for same initial setup and input parameters: in one run the phonon spectrum shows a stabilized structure with no imaginary frequencies, while in another run imaginary frequencies appear .Is it expected that the renormalized phonon frequencies or phonon bands from two separate runs would be exactly the same? Or is some variation anticipated due to the stochastic nature of MD and sampling of anharmonic effects? Any advice or experiences related to understand such differences in phonon calculations with MLFF MD would be greatly appreciated.

[manuel_engel1]

Hello and welcome to the VASP forum,

It is not expected that you get exactly the same results. When you run an MD with MLFF, the outcome can be different depending on various parameters and sometimes even the hardware. Depending on your system and your calculational parameters, the dynamics and final state of your simulation can vary significantly. It could be due to the physical nature of the material, for example, capturing large anharmonicities or instabilities close to a phase transition. However, it might also be that your computational setup does not have enough accuracy, does not converge or does not describe the system well.

I highly encourage you to play around with the computational parameters and study your system to understand why these differences exist. Let me know if you have any specific questions.

[mitanshi_gupta]

For my system their exists various polymorphs, I want to stabilise the phonon spectrum for one of the polymorph around 550K for that i am training it at 750K and then doing production run at various temperatures so i am getting stablised phonon (no imaginary frequencies)at 610K and above and below it imaginary modes are there, but if using that trained forcefield i am training it again at 750K for more 20000 no. of steps then i am not getting stabilised phonon at 610K but at both temp. 600K and 670K and between this range there are imaginary modes. So i am confused which temperature is my transition temparature. And also can i use the final MLFFN to further train at higher temperatures?

[manuel_engel1]

Can I ask you how you extract the phonon frequencies from these calculations? Negative phonon frequencies are not always an indicator of instabilities or phase transitions. In particular, phonons computed via perturbation theory or finite differences (IBRION=6 or 8) are really only well defined at equilibrium and not at high temperature.

[mitanshi_gupta]

I am using dynaphopy which is giving me renormalised phonon for higher temperatures. In my understanding, i am thinking that i should get a renormalised stabilised phonons (no imaginary frequencies) at a nearby temperature where the phase transition is happening experimentally which is 550K, to achieve that i am doing training for 750K and then doing production run from 400K-700K and in between(610K) i am getting stabilised phonons. Am i right here in doing this?

Also can i use the force field for further training at more higher temperatures?

If i am doing so, and training it at 800K then the temperature at which i am getting stabilised phonon is different from trained at 750K. So how to decide which temperature for training is correct or more accurate?

[mitanshi_gupta]

I'm attaching my INCAR and ML_LOGFILE for training and production.

Also i am using 222 supercell which is 40 atoms for training So i should do prediction for higher supercell(333)?
How many steps(NSW) are good enough for considering a forcefield trained and also for doing production. Because as the no. of steps is increasing in production run, i am getting different results.

[manuel_engel1]

I have to admit that I'm out of my depth here at this point. Maybe we can still get some input from the machine-learning team. Otherwise, there might be some people in the community who can contribute to this question.

[mitanshi_gupta]

Hey can you please confer this with your machine learning molecular dynamics team ? As if they have not responded.

[andreas.singraber]

Hello!

As far as I understand from your messages and the files you sent, you performed training runs only at 750K, is that correct? In your data I did not see a gradual heating up with TEBEG < TEEND but just both set to 750K. Because the heating will then happen abruptly, the training data will mostly contain structures representative of the distribution at 750K with hardly any structures corresponding to lower temperatures. However, you mentioned that you want to apply the force field at another temperature range (400K - 700K, in particular 610K). Then your force field will mostly extrapolate in these regions and results will not be reliable.

I suggest to carry out training runs with gradually increasing temperature to sample the temperature range uniformly. At specific temperatures of interest you may even want to carry out additional (continuation) training runs to improve the force field's reliability there. You may need to bounce forth and back between phonon calculations and retraining until you see that results converge. You mentioned a transition region at 550K.. if you intend to have a reliable force field there you should be particularly careful to sample the phase transition. Maybe a good starting point on how a training protocol could look like is described in this paper (which also uses machine-learned force fields of VASP):

https://ris.utwente.nl/ws/portalfiles/p ... attice.pdf (https://doi.org/10.1103/PhysRevB.105.024302)

Please understand that we cannot guide you through the entire process of generating reliable data for your system. However, there are helpful hints and tricks described at our "Best practices" Wiki page here:

https://vasp.at/wiki/Best_practices_for ... rce_fields

All the best,
Andreas Singraber