My Community

Dear all,

I'm trying to generate MLFF for amorphous system.

I trained the amorphous SiO2 from 2000K to room temperature with NVT ensemble. After I trained the model, I tested it to reproduce amorphous SiO2. In many cases, it gave reasonable structures, but sometimes there exist O-O or Si-Si overlaps. I think it is because MLFF doesn't have a specific functional form, thus it didn't reproduce strong repulsion b/w atoms at a very short distance, which is in below 0.5 A.

Thus, I think I need to additionally train the system with O2 or Si system. But, I still not sure it can reproduce strong repulsion at a very short distance. Are there any good way to train it?

If any one of you have experiences, please let me know.

Regards,

I think it's an indication that your force field is extrapolating.

What is the target of the test structures? Your training temperature (the end of the temperature ramp) should be higher than your training structures.
You also mentioned you trained from 2000K to room temperature. I hope you didn't do a cooling run that will not work with the defaults of the machine learning. You always must use a heating run.
I would also suggest to use NpT ensemble since due to the volume fluctuations the force field gets more stable. Please see our best practices site:
https://www.vasp.at/wiki/index.php/Best ... rce_fields
Instabilities can sometimes also be cured by just running the on-the-fly sampling simply much longer (let's say another 100 ps).

Thanks for your comment!
I agree on your comments, extrapolating. Thus, I wanted to knwo how can I include energies or forces for the very short distances.

I trained the structures as follows,
1. Melt at 2000K for 20 ps
2. Cooling from 2000K to 300K for 20 ps
3. Relax at 300K for 10 ps

Since I will use MLFF for melting and cooling, I used the cooling process during training, but it may not right based on your comments. Let me try the heating and melting cycle only.

Regarding the NpT ensemble, the random structure often fall into the wrong density, thus I only applied NVT ensemble. Let me try it as well.

Thanks!