Training a DeePMD model on biased trajectories?

[kuntalg97]

Hi,

I'm trying to model a system with DeePMD where my initial dataset comprises trajectories from umbrella sampling. I used LAMMPS and PLUMED for generating the umbrella sampled trajectories. Now, I wish to train a DeePMD model on these trajectories so that I can use the resultant neural network (in LAMMPS with PLUMED) for reproducing the atomistic PMF of the phenomenon that I'm studying.

What are the best ways to train on biased trajectories? One way to do so might be to remove the biasing forces from the trajectories from each window? What I did was rerun the trajectories through LAMMPS (using the 'rerun' command) but with the original force fields without any bias, to generate forces corresponding to the unbiased force field. I then trained a DeePMD model on these reprocessed trajectories (with the same coordinates as the original but different forces) and ran the resultant model with LAMMPS and PLUMED, but my free energy curves were horribly messed up.

So my question is - how do I train DeePMD on biased trajectories? How do I re-process the biased trajectories such that the above problem can be solved?

Thanks!