My Community

I am training a force field of AgF, NaCl and NaI cells, using a temperature ramp from 50 to 500 K. The three cells have the same structure, but different lattice parameters, which I relaxed beforehand.

I use exactly the same settings for all three systems, but during training, only the AgF behaves as it should, i.e. the temperature slowly rises, and the force field is updated with ab initio steps after some iterations. For the NaCl and NaI, the temperature stays roughly the same. Also, the Bayesian error estimation remains very low and the force field is almost never updated.

I am uploading the files for my NaCl simulation with PBE-D3 functional, except for the OUTCAR, which is 15 MB. The images show the monitoring of the Ab initio and ML properties for all three systems. The other functionals that I have tried behave identically.

What am I doing wrong? Is this only related to the missing temperature ramp, or do I need to change the learning threshold as well?

Hi, I will try to figure this out with you.

First of all, I notice that you are setting MDALGO = 3 but no value for LANGEVIN_GAMMA. The default is LANGEVIN_GAMMA = NTYP*0, hence there is no coupling to the heat bath. I am rather puzzled as to how you can get the AgF result like that. Please have a look at the different thermostats and let me know if the explanations are clear enough!

Regarding the large OUTCAR file, please check out ML_OUTBLOCK to reduce the output to what you need.

Thank you for the reply. I have added the LANGEVIN_GAMMA and LANGEVIN_GAMMA_L flags to 10 for each species. Let's see how it works.

How can I make an educated guess for the proper magnitude of the friction coefficient? I am using 10 ps^-1 now only because I found this number in a different thread, but I don't know if it is appropriate for my system.

How did it go?

It is not easy to make an educated guess. In practice, one needs to run a couple of test values for LANGEVIN_GAMMA. The optimal setting is also influenced by the value for POTIM.