compute_extrapolation

This keyword is used to compute the extrapolation grade of structures in an NEP potential.

The extrapolation grade gamma can be considered as the uncertainty of a structure relative to the training set.

A structure with large gamma tends to have higher energy and force errors.

Similiar methods have been applied to MTP ([Podryabinkin2023]) and ACE ([Lysogorskiy2023]). You can refer to their papers for more details.

Before computing gamma, you need to obtain an active set from your training set. There are some Python scripts to do it <https://github.com/psn417/nep_active>.

There are also some Python scripts to perform active learning automatically <https://github.com/psn417/nep_maker>.

Syntax

This keyword is used as follows:

compute_extrapolation asi_file <asi_file> gamma_low <gamma_low> gamma_high <gamma_high> check_interval <check_interval> dump_interval <dump_interval>

asi_file is the name of the Active Set Inversion (ASI) file. This file is generate by the Python script in <https://github.com/psn417/nep_active>.

gamma_low: Only if the max gamma value of a structure exceeds gamma_low, then the structure will be dumped into extrapolation_dump.xyz file. The default value is 0.

gamma_high: If the max gamma value of a structure exceeds gamma_high, then the simulation will stop. The default value is very large so it will never stop.

check_interval: Since calculating gamma value is slow, you can check the gamma value every check_interval steps. The default value is 1 (check every step).

dump_interval: You can set the minimum interval between dumps to dump_interval steps. The default value is 1.

Example

compute_extrapolation asi_file active_set.asi gamma_low 5 gamma_high 10 check_interval 10 dump_interval 10

This means that the structures with max gamma between 5-10 will be dumped. The gamma value will be checked every 10 steps.