Machine Learning Potential

[qingyu_wang]

Hello,professor
Thank you very much for your reply. I have now completed the optimization of the initial structure. Subsequently, based on this optimized configuration, I performed Ab Initio Molecular Dynamics (AIMD) calculations with the parameter ML_MODE=select, using the ML_AB file obtained as mentioned above. Following this, I conducted training with the parameter ML_MODE=train on the aforementioned optimized structure. Could you please advise whether this approach is methodologically correct? Furthermore, I have checked the variations of energy and forces, and they all align with my predefined requirements. In addition, all other settings have been configured with reference to the suggestions you provided.
Best wishes!

[max_liebetreu]

Hello,

There's no need to run ML_MODE=select on the initial structure. Please check the wiki entry for specifics on the select option. As long as you keep copying the ML_ABN file to ML_AB for each step of your training (pure -> 1% -> 5% -> 10%), you shouldn't need to do select at all.

The important thing is that you:

1. relax your structure,
2. train,
3. copy your ML_ABN file to ML_AB for the next structure,
4. back to 1. for the next structure.

If you follow these steps, you don't need to combine the different ML_ABN files at the end of your training, and you also don't need to reselect with ML_MODE=select. The final ML_ABN will already have all that information - you just need to move it to the folder where you want to run your next computations, and rename it to ML_AB. Additionally, for simply using the final force field (without training), you also need ML_FFN as ML_FF with ML_MODE=run.

I hope that clears up any remaining confusion!
Good luck, and best regards,

[qingyu_wang]

Dear Professor,
Thank you very much!
Best wishes!