Using DeePMD for training on single point QM optimized structures and then doing dynamics in LAMMPS

[kuntalg97]

Hi,

I have a set of structures generated from CP2K with QM minimization. I wish to train a DeePMD model on these structures, and then deploy these in a molecular dynamics simulation in LAMMPS. In other words, I want my DeePMD model to accurately learn the PES along some reaction coordinate and then use that as a force field in LAMMPS.

My case is very similar to this - #3416

I've already prepared my data (I have experience with training DeePMD models on MD trajectories). I also had the 'nopbc' file in my training/validation directories. However, the trained DeePMD models crash almost instantly while running on LAMMPS. Is my training set too small or is it something else? Does the dimensions in box.raw matter? Can a DeePMD model trained on non-periodic conditions be directly used for MD with periodic conditions?

How do I proceed with this?

Thanks!

[wanghan-iapcm]

In general, it is impossible to answer if a model trained by nopbc data is applicable in MD simulations under pbc.

The critical consideration is not the configuration of the boundary conditions in the training data, but rather whether the local atomic configurations are adequately sampled by the training dataset. In your specific case, it is essential that the reaction events defined along the reaction coordinate are thoroughly sampled. To achieve this, the use of concurrent learning is recommended.

concurrent learning method
https://doi.org/10.1103/PhysRevMaterials.3.023804
and implemented by dpgen
https://doi.org/10.1016/j.cpc.2020.107206

[kuntalg97]

Thanks! I'll see what I can do.