My Community

Dear VASP developers,

To train an on-the-fly machine learning force field using VASP, we need to perform at least one simulation with ML_MODE=TRAIN, and then refit a ML_FFN file based on the ML_AB file. However, due to memory limitations, a single on-the-fly training from scratch, especially for systems with three or more elements, cannot run for a long time, and tends to be slower and slower. Continuing on-the-fly training from an existing training database can also be very slow. On the other hand, it seems feasible to perform a series of on-the-fly training runs (from scratch) with different initial structures, and then merge the resulting ML_AB files. This approach could also be beneficial for combining ML_AB files from different systems, such as water and sodium chloride aqueous solutions.

The VASP wiki of ML_AB provides detailed instructions for merging ML_AB files, but the process requires numerous modifications and careful attention. I am wondering if the VASP team could provide an official script to facilitate this merging process, which would greatly improve compatibility and usability. Thank you for your time and consideration.

Sincerely,
Tian Hua

Dear xiancai,

We currently are working on a major rewrite of our machine learning framework. Merging two ML_AB files will be a feature included in our new machine learning code when released.

All the best Jonathan