I'm trying to apply the active learning procedure, which was explained in issues #62 and #158.
However, i noticed that the potential energy values produced from lammps interface and from nnp-comp2 are quite different.
I'll drop the files in a zip and quote the filenames below for ease of understanding.
I started the procedure by producing trajectories from a trained nnp through n2p2/lammps interface. The energies.data file contain the potential energy values for each step of simulation. So i converted the lammps output file in a input.data file so that i could use the nnp-comp2 tool.
However, when i ran the tool, the prediction values in diff.out were quite different comparing with the output lammps values ("E_NNP1" and "E_NNP2" columns). The model 2 is basically the same of the model present in nnp-data-1 folder (i just changed the "seed" keyword value).
The structure that i used to make that procedure was in database used to train the nnp model, so i think the potential energy values shouldn't be so different. The two hypotheses that i considered were that either the number of structures in database is still low or I haven't entered any configuration in pair_style.
I'm trying to apply the active learning procedure, which was explained in issues #62 and #158. However, i noticed that the potential energy values produced from lammps interface and from nnp-comp2 are quite different.
I'll drop the files in a zip and quote the filenames below for ease of understanding.
I started the procedure by producing trajectories from a trained nnp through n2p2/lammps interface. The energies.data file contain the potential energy values for each step of simulation. So i converted the lammps output file in a input.data file so that i could use the nnp-comp2 tool.
However, when i ran the tool, the prediction values in diff.out were quite different comparing with the output lammps values ("E_NNP1" and "E_NNP2" columns). The model 2 is basically the same of the model present in nnp-data-1 folder (i just changed the "seed" keyword value).
The structure that i used to make that procedure was in database used to train the nnp model, so i think the potential energy values shouldn't be so different. The two hypotheses that i considered were that either the number of structures in database is still low or I haven't entered any configuration in pair_style.
nnp.zip
Thanks in advance for any help.