Hello,
In my recent attempts at constructing MLFFs, one issue I've had is the unphysical clustering of atoms on top of each other (<0.2 angstroms apart) during an MLFF production run, an issue I haven't had in either the training data used for the MLFF or in ab initio data for the same structure which is a mixed fluid. My question is if this is a common occurrence with multi-element fluid systems, and if there may be some intuition how to address it. My first thought was maybe this means there are not enough training structures or the quality of the training data is poor, but even so I don't understand how this could cause clustering of atoms in the MLFF prediction. Any thoughts on this are appreciated!
Thank you,
Dylan
Atoms clustering together in MLFF prediction run
Moderators: Global Moderator, Moderator
-
dylan_durkee
- Newbie
- Posts: 10
- Joined: Mon Feb 20, 2023 3:46 pm
-
andreas.singraber
- Global Moderator
- Posts: 236
- Joined: Mon Apr 26, 2021 7:40 am
Re: Atoms clustering together in MLFF prediction run
Hello Dylan,
I would agree that this is likely due to insufficient training data and/or untrained temperature or pressure conditions in the production run. During training it is necessary to establish that a strong repelling force will result from atoms coming very close to each other. Most likely you already have some atomic environments in your data set which imprint this repelling force into the MLFF. However, the smallest distance between atoms in your data set will implicitly define the limit where extrapolation begins. In this regime we cannot guarantee that the functional form of the MLFF for even closer atoms will still result in a repelling force, it could even be the other way round and an attractive force "glues" the atoms together.
Now, your production run is probably using a larger system size or at least more time steps than the training run. Hence, there is a higher chance that two atoms are pushed very close together and the MLFF goes into extrapolation. In your case this led to an even qualitatively wrong behavior as the atoms were not repelled but rather attracted after crossing the barrier.
To avoid this problematic behavior in the future I suggest to collect more training data at higher temperatures and/or higher pressures to include more samples of atomic environments with very close atoms.
Best,
Andreas Singraber
I would agree that this is likely due to insufficient training data and/or untrained temperature or pressure conditions in the production run. During training it is necessary to establish that a strong repelling force will result from atoms coming very close to each other. Most likely you already have some atomic environments in your data set which imprint this repelling force into the MLFF. However, the smallest distance between atoms in your data set will implicitly define the limit where extrapolation begins. In this regime we cannot guarantee that the functional form of the MLFF for even closer atoms will still result in a repelling force, it could even be the other way round and an attractive force "glues" the atoms together.
Now, your production run is probably using a larger system size or at least more time steps than the training run. Hence, there is a higher chance that two atoms are pushed very close together and the MLFF goes into extrapolation. In your case this led to an even qualitatively wrong behavior as the atoms were not repelled but rather attracted after crossing the barrier.
To avoid this problematic behavior in the future I suggest to collect more training data at higher temperatures and/or higher pressures to include more samples of atomic environments with very close atoms.
Best,
Andreas Singraber
-
dylan_durkee
- Newbie
- Posts: 10
- Joined: Mon Feb 20, 2023 3:46 pm
Re: Atoms clustering together in MLFF prediction run
Hi Andreas,
Thank you for your helpful response, it gives me some ideas moving forward!
Thank you for your helpful response, it gives me some ideas moving forward!