Error estimate during MD simulation

[singraber]

With multiple networks (different random seed at training) calculating the same total energy at each timestep, we could compute a spread (min/max-difference or standard error) of energies which indicates missing information in the training data.

Since no additional symmetry functions need to be calculated the performance will not be affected much.

[singraber]

This is similar to what the tool nnp-comp2 tool does, but it would be nice to have it on-the-fly and with more networks.

[marian-code]

This would be really useful feature to have.
Also a small suggestion, it would be nice if the code could output the structures not present in the dataset (based on some user defined criterion) to a file so one could easily take and incorporate them into training dataset.

[singraber]

This still has to be implemented.. however, some prerequisites for such an enhancement (more NNs, refactored NN) are currently WIP: #64 and #70

[ZengZezhu]

Hi,

I would like to know whether we can run the MD simulations based on multiple-NNPs (just like an average of multiple-NNPs) in lammps now? Thanks.

Cheers,
zezhu

[singraber]

Hey,

there is an unfinished pull request #92 which I still need to test. So unfortunately not yet, sorry...

Best,
Andreas

[ZengZezhu]

OK...
Yes, I see C-NNPs (generated by AML package: https://github.com/MarsalekGroup/aml) can run in the CP2K now, but for lammps, it is still blank. I am looking forward to this part. :) Thanks.