Sparsification of local reference configurations

[julien_steffen]

Dear VASP community,

we are currently trying to learn a system containing Ga and H atoms within a ML-FF calculation. To this reason, we have performed four different calculations with four different starting configurations regarding the H.

After that we have combined the four corresponding ML_AB files to one big ML_AB file by selecting 1000 reference configurations for each element.

However, when we are trying to do a REFIT calculation to get a ML_FFN file from the ML_AB file (the actual force field) and to use the FAST mode of VASP ML_FF production runs the number of local reference configurations is reduced/sparsificated drastically.

According to the VASP Wiki this can be limited by setting the ML_EPS_LOW keyword to smaller values. However, this has not worked sufficiently for our system. The sparsification is still too high. Therefore, the Force Field gets instable by forming H3 clusters after a while in the production run.

Is there a way to completely switch off the sparsification of local reference configurations during a REFIT calculation that we can use the prior selected 1000 basis functions for the kernel?

Since the size of files is limited for the VASP Forum, please find the used input and output files at:

https://faubox.rrze.uni-erlangen.de/get ... Wohai43rh/

Best regards

[alexey.tal]

Hi Julien,

However, when we are trying to do a REFIT calculation to get a ML_FFN file from the ML_AB file (the actual force field) and to use the FAST mode of VASP ML_FF production runs the number of local reference configurations is reduced/sparsificated drastically.

The local reference configurations which are not sufficiently different, i.e., yield the same eigenvalues, are sparsificated. You can reduce the threshold for the reduction with ML_EPS_LOW=10E-14.

Is there a way to completely switch off the sparsification of local reference configurations during a REFIT calculation that we can use the prior selected 1000 basis functions for the kernel?

At the moment there is no way to switch off the sparsification, you can only reduce the threshold. But we will consider introducing this option in the future.