Information and Capabilities of Pretrained MTPs in Force Field Library for MD

[msingh96]

Hello QuantumATK users, this is my first post in the forum. I have started using QuantumATK for the MTP functionality it provides as advertised here: https://www.synopsys.com/manufacturing/quantumatk/atomistic-simulation-products/machine-learned-force-fields.html.

One exciting thing I found was the library of pretrained MTPs made available for users: https://docs.quantumatk.com/manual/ForceField.html#pretrained-moment-tensor-potential-mtp-parameter-sets

However I am struggling to use these effectively. I am running into two errors.
1. the nose-hoover thermostats seem to fail/atoms seem to explode for certain interfaces
2. atoms leave simulation regime and MD simulation stalls.

I am wondering if there are any examples or  information about the training data (structures) for these pretrained potentials.

Regards and Thanks, Mayur

[Anders Blom]

If you go outside the domain where the potentials are trained you can easily see such effects, yes. Which MTP in particular are you trying to use?

[msingh96]

Hi Anders, I've been trying to figure this out on my own before posting again.

I am trying to use the Si/TiSi₂ potential for MD simulations. I have had modest success with small structures, but the most frequent error is atoms leaving the simulation cell. this error causes the MD simulation to stall and get stuck. my instinct is an issue with the structures I provided, I used the interface builder to make an initial structure with Si [100] surface and TiSi[001] surface but I have to manually modify the distance between the surfaces using shift surface. I want to equilibrate the surface distance suing NPT but that's not working either.

Are there any example geometries from the training data for this potential to see where I'm going wrong?

[msingh96]

Hello. I think I figured it out. the distance between surfaces in the initial geometry was too close, causing large forces to explode and break most structures.

[Anders Blom]

Yeah MTP can be particularly sensitive to this, forces can diverge in such geometries. We have been experimenting with adding a short-range repulsive potential (ZBL) but it doesn't always work as well as desired. Good that you solved it!