MLFF procedure and error of the version of vaspml when running the trained step

[manoj_wijesingha]

Hello Colleagues,
When I was running trained MLFF by the latest version of VASP.6.5.0, this error occurred; "The given force field file is too new 0.2.4), this vaspml distribution only supportsfiles up to version 0.2.3." These tags are included in the INCAR file, ML_ISTART = 2 and NSW = 1000. Could please support to solve this.

Next, I am very new to use the machine learning and still learning the process. So, I need to use machine learning for my VASP simulations, and I currently use VASP.6.5.0 and I am able use MLFF in VASP. Since I am lacking the knowledge about using machine learning, could you please guide, suggest, and provide the steps by step to use machine learning simulation. For example, to obtain the adsorption energy of CO2 (or any example) on 2D or 3D materials, and I need to apply machine learning for this simulation by MLFF in VASP. When I use the AVSP.6.5.0 to simulate by machine learning, should I use DeepMD-kit, AMP or NN for this simulation. Finally, I need to analyze and compare the density functional theory (DFT) and machine learning values. Could you explain this example from the beginning with DFT and then MLFF, and this is really helpful and appreciate. 
Thank you

[ferenc_karsai]

Can you please specify what were you doing?
This sounds a little bit like you were trying to run an ML_FF file from VASP.6.5.0 with an earlier version of VASP.
Please do the training (ML_MODE=train), refitting (ML_MODE=refit) and production (ML_MODE=run) with the same version of VASP.

VASP is perfectly feasable to run all calculations neccessary with and without machine learning to obtain adsorption energies.

A tutorial for molecular dynamics (MD) and machine learning force fields (MLFF) in VASP is given here:
https://www.vasp.at/tutorials/latest/md/

The how to for machine learning force fields is here:
https://www.vasp.at/wiki/index.php/Mach ... ns:_Basics

Please also look into important hints for both MLFF with MD here:
https://www.vasp.at/wiki/index.php/Best ... rce_fields

A description of the procedure how to obtain adsorption energies using MLFF would be out of the scope of this forum, but here is a publication that you can consider:
https://doi.org/10.1103/PhysRevLett.130.078001

[andreas.singraber]

Hello!

The first issue was already discussed in this forum, there is indeed a bug in the new VASPml library. please see this post here:

https://www.vasp.at/forum/viewtopic.php?p=30485#p30485

Unfortunately, last week we were not able to release the next VASP version where this bug is already fixed... but it should be a matter of days now...

Please be aware that the VASPml library is in an experimental stage and some bugs are still possible. We list the ones already found here:

https://www.vasp.at/wiki/index.php/VASP ... strictions

Maybe you did not intentionally compile with this experimental library (-Dlibvaspml). If you do not want to use it, you can always disable it with ML_LIB = .FALSE. in the INCAR file.

All the best,
Andreas Singraber

[manoj_wijesingha]

Hi Karsai,
Thank you and really appreciate your response. Your attachment is also very informative. So, I need to study and apply MLFF for my simulations like these publications, "https://doi.org/10.1039/d3sc05612k and https://doi.org/10.1021/acsnano.2c11102".

I used only the version of VASP.6.5.0 but still generating. I could successfully run MD, and then ML_MODE=train, and ML_ISTART = 1, ML_MODE = run. But when running with this key word in the INCAR file, ML_ISTART = 2, this error happens. Thank you.

[ferenc_karsai]

As Andreas has written above you have compiled with -Dlibvaspml and you can fix your issue by using ML_LIB = .FALSE. in the INCAR file.

Another thing is please don't use the deprecated tag ML_ISTART. The tag ML_MODE selects all necessary modes.

[manoj_wijesingha]

Thank you all and really appreciate it.