How can I include trajectories?

[EYErbil]

Hello, I have trained my model,

but what I realized is, it has problems when I relax a perturbed structure, it basically converges, but in a very stupid way,

I have trained my model with single point snapshots, meaning that it doesnt know how to proceed to next geometry actually,
do you think I can include relaxation trajectories of my DFT calculations from before? would that help?

[wanghan-iapcm]

Yes, including DFT relaxations will help the generalization in relaxation tasks.

A better way is to use the concurrent learning that only collects and labels (by DFT) critical configurations along the relaxation trajectories. please refer to https://journals.aps.org/prb/abstract/10.1103/PhysRevB.109.094117 for more details

[EYErbil]

Oh, then the calculations that have trajectories, should have those in the output file of cp2k, right? I dont need to do anything extra, converting the out file with cp2kdata plugin would be enough for the model to understand the trajectories

[wanghan-iapcm]

The trajectories are supposed to be convert to format "deepmd/npy" that can be understood by the deepmd-kit.
you may find examples in the tutorial https://docs.deepmodeling.com/projects/deepmd/en/r2/getting-started/quick_start.html#data-preparation

[EYErbil]

Yes, but then, the trajectory files created by cp2k, how do I convert them to NPY, I mean, trajectory files are only .xyz, they dont have any energy attached to them, only positions, or do I do this wih only the output files of relaxation procedures in cp2k?
os.system(f"dpdata {output_file} -i cp2k/output -o deepmd/npy -O {data_dir}")
system = dpdata.LabeledSystem(file_name=data_dir, fmt='deepmd/npy')

[EYErbil]

In this code, I convert .out files to .npy, as labeled systems, and append to multi systems, all output files are single point, so that means no trajectory is present,

[wanghan-iapcm]

What I mean is that every "single point" on your optimization trajectory, with labels, can be used for training data.