search

brucefan1983 / GPUMD

Graphics Processing Units Molecular Dynamics
https://gpumd.org/dev
GNU General Public License v3.0
466 stars 116 forks source link

consider an active learning scheme in NEP #130

Closed brucefan1983 closed 2 years ago

brucefan1983 commented 2 years ago

This is a must-have sooner or later. Might be contributed by someone.

brucefan1983 commented 2 years ago

Now one can run NEP-MD with small boxes. With this, one can easily do active learning using, e.g., the ensemble method:

(1) Train a few (say, 5) independent NEP potentials for your system. (2) Run some MD simulations with one or more NEP potentials, saving the trajectories. (3) Compute the properties (energy, force, and virial) for the saved trajectories using all the NEP potentials. (4) Determine which configurations in the trajectories need to be further examined based on the deviations of the properties (energy, force, and virial) predicted by the different NEP potentials. 5 Select a number of configurations and add them into the original training data set and got to step (1).

There are other ways of doing active learning of course, but it is not my interest to couple GPUMD with any DFT code to realize one. There are just too many possibilities.

brucefan1983 commented 2 years ago

Now there is a good active-learning scheme for NEP, see PyNEP: https://github.com/bigd4/PyNEP