Using Allegro to train a model on single point QM optimized configurations (non-periodic conditions)

[kuntalg97]

Hi,

I have a set of structures generated from CP2K with QM minimization. I wish to train an Allegro model on these structures, and then deploy these in a molecular dynamics simulation in LAMMPS.

The few tutorials that I've come across train on molecular dynamics data (such as this) however I am specifically looking for a way to learn the PES from single point calculations only. This would also imply that my training set has no periodic boundary conditions.

How do I proceed with this? As in, how do I explicitly account for non-periodic boundary conditions, as my training set is not technically a 'trajectory'.

Thanks!

[cw-tan]

Could you please define the "single point calculations" that you're referring to? In principle molecular dynamics data (if you perform the full SCF at each step and not say Car–Parrinello MD), or relaxation trajectories can be thought of as sets of "single point calculations" if you define "single point calculations" as full DFT SCFs. I'd also say that periodic boundary conditions are independent with respect to whether the frame was from a "single point calculation" or "trajectory" in general. Please clarify these points for us to give you better advice. Thanks.

[kuntalg97]

Sure. Say I have a simple SN2-type reaction in vacuum: CH3Cl with a bromide ion. I'm constructing a set of structures such that the difference between the C-Cl and C-Br distances is within a particular range of values (consider this distance difference to be a reaction coordinate). I'm minimizing each structure at some level of QM theory (DFT/full CI, doesn't really matter). Now I wish to train an Allegro model on this set of structures, the goal being to have the NN model accurately reproduce the PES along the aforementioned reaction coordinate. I'm generating the forces on these by re-running these structures through CP2K.

[cw-tan]

Is it possible to use ase.io.read to parse your output files and stack them into a single .extxyz (extended xyz) file? see https://wiki.fysik.dtu.dk/ase/ase/io/io.html. There's a CP2K ASE calculator, so hopefully it's possible. And hopefully the way ASE parses your output files will respect whether your simulations use periodic boundary conditions or not.

[kuntalg97]

Thank you, I'll take a look into this and see if this helps.