Training DPLR in Active learning process

[karvs25]

Hi All,

I have a doubt that I did not train the Deep Wannier model using an active learning process, since I dont know how to use DPGEN for calculating Wannier centres. Is it reliable to use the active learning process to train DPLR model with the dipole model created using passive learning? Does it make sense since the data will be different from the DPLR model and the Deep Wannier model?

Best
Suresh

[sgindeed]

You can train a DPLR model using an active-learned DP model together with a Deep Wannier model trained through passive learning, but it’s only reliable if the two datasets cover similar configuration space. The main risk is that DP-GEN will sample new environments that your passive DW model never saw, leading to inaccurate Wannier centres and therefore incorrect long-range electrostatics. To stay safe, you should check whether the active-learning configurations fall within the domain of your DW training data; if not, compute Wannier centres for a representative subset of the new structures and fine-tune the DW model. As long as you ensure this consistency and validate dipoles/electrostatic behaviour on held-out configurations, using passive-trained DW with active-trained DP in a DPLR workflow is reasonable.

[karvs25]

Thank you for your reply, @sgindeed.