search

mir-group / flare

An open-source Python package for creating fast and accurate interatomic potentials.
https://mir-group.github.io/flare
MIT License
278 stars 64 forks source link

Time of Update GP (s) increases a lot as more DFT calls #372

Open YukeLi-chem opened 10 months ago

YukeLi-chem commented 10 months ago

Hi Flare developer,

Offline training is doing. With more DFT calls, the Time of Update GP (s) increases a lot shown in figures. And the train_hyps fixed as 'train_hyps = [20, 30]' to prevent the hyperparameters optimization.

train

Best, Li Yuke

anjohan commented 9 months ago

Hi,

Sorry for the late reply!

Since GPs require matrix inversion, they scale cubically in the number of training points. For FLARE, hundreds of structures is considered a fairly large training set. Make sure that you only add the structures and environments actually needed (as determined by the uncertainty).

You may be able to improve the performance by playing around with OpenMP and BLAS parallelization settings. Generally, MKL gives the best performance.