ML MB MIN: Difference between revisions

From VASP Wiki
No edit summary
No edit summary
Line 4: Line 4:
Description: {{TAG|ML_MB_MIN}} sets the minimum number of local reference configurations required for generation of a machine-learned force field. {{NB|mind|Available as of {{VASP}} 6.4.3. Previous versions do not support this tag but act as if its default was {{TAGO|ML_MB_MIN|2}}.}}
Description: {{TAG|ML_MB_MIN}} sets the minimum number of local reference configurations required for generation of a machine-learned force field. {{NB|mind|Available as of {{VASP}} 6.4.3. Previous versions do not support this tag but act as if its default was {{TAGO|ML_MB_MIN|2}}.}}
----
----
Machine-learned force fields based on the kernel method require a reasonable amount of reference atomic environments in order to provide meaningful predictions. Typically, the on-the-fly algorithm in {{VASP}} collects these local reference configurations (also called kernel basis functions) during an MD simulation ({{TAGO|ML_MODE|train}}). However, at the start of the trajectory usually only very few local reference configurations are available. How many depends on the system size, symmetry and number of atoms (per type).
Machine-learned force fields based on the kernel method require a reasonable amount of reference atomic environments in order to provide meaningful predictions. Typically, the on-the-fly algorithm in {{VASP}} collects these local reference configurations (also called kernel basis functions) during an MD simulation ({{TAGO|ML_MODE|train}}). However, at the start of the trajectory usually only very few local reference configurations are available. How many depends on the system size, symmetry and number of atoms (per type). This tag controls how many local reference configurations are required '''for each atom type''' before training of a force field is allowed. Setting a higher value may yield a more robust initial force field at the cost of  more ''ab initio'' calculations during the first few MD steps.


If a training is scheduled in the current step, i.e., the <code>STATUS</code> line in the {{FILE|ML_LOGFILE}} shows <code>learning</code> or <code>critical</code>, but not enough atomic environments were collected, a text message will be emitted in the <code>MSG</code> line. Here is an example:
--------------------------------------------------------------------------------
STATUS                  1  critical  4      T      F        0        1
LCONF                  1 Ni        0        4
SPRSC                  1        1        1 Ni        4        1
MSG                    1      info Number of local reference configurations after sparsification below ML_MB_MIN, skipping training.
BEEF                    1  1.00000000E-06  3.46410162E-02  2.00000000E-02  2.00000000E-03  0.00000000E+00  0.00000000E+00
--------------------------------------------------------------------------------
STATUS                  2  critical  4      T      F        0        2
LCONF                  2 Ni        1        5
SPRSC                  2        2        2 Ni        5        2
MSG                    2      info Number of local reference configurations after sparsification below ML_MB_MIN, skipping training.
BEEF                    2  1.00000000E-06  3.46410162E-02  2.00000000E-02  2.00000000E-03  0.00000000E+00  0.00000000E+00
--------------------------------------------------------------------------------
STATUS                  3  critical  4      T      F        0        3
LCONF                  3 Ni        2        6
SPRSC                  3        3        3 Ni        6        3
REGR                    3    1    1  6.42271768E+01  1.36891529E+00  6.43938951E-14  3.76458019E+01
REGR                    3    1    2  7.43557292E+01  1.35099932E+00  5.48943502E-14  3.71479653E+01
REGR                    3    1    3  7.81437752E+01  1.34829112E+00  5.21286212E-14  3.70727138E+01
REGRF                  3    1    4  7.95268527E+01  1.34739777E+00  5.11880969E-14  3.70478685E+01    2.31498305E+15  3.30988608E+11
NDESC                  3        12 Ni      113
NDESC_SIC              3 Ni      113
STDAB                  3  1.41895828E-04  3.22944034E-02  9.66711031E-02
ERR                    3  1.45620371E-04  3.32055814E-02  9.84035846E-02
CFE                    3  0.00000000E+00  0.00000000E+00  0.00000000E+00
LASTE                  3  1.99936842E-04  5.14296619E-02  1.37823063E-01
BEE                    3  1.00000000E-06  3.46410162E-02  2.00000000E-02  2.00000000E-03  0.00000000E+00  0.00000000E+00
BEEF                    3  2.37926034E-05  2.21630107E-03  8.49321832E-04  2.00000000E-03  5.19041973E-02  3.66970792E-02
--------------------------------------------------------------------------------
...
Here, in the first two steps the {{TAG|ML_MB_MIN}} threshold prevents training of an MLFF, even if the Bayesian error estimation signals a <code>critical</code> step. The force field is then only generated in the third step, after the minimum number of three local reference configurations have been collected.


== Related tags and articles==
== Related tags and articles==

Revision as of 09:39, 19 February 2024

ML_MB_MIN = [integer]
Default: ML_MB_MIN = 3 

Description: ML_MB_MIN sets the minimum number of local reference configurations required for generation of a machine-learned force field.

Mind: Available as of VASP 6.4.3. Previous versions do not support this tag but act as if its default was ML_MB_MIN = 2.

Machine-learned force fields based on the kernel method require a reasonable amount of reference atomic environments in order to provide meaningful predictions. Typically, the on-the-fly algorithm in VASP collects these local reference configurations (also called kernel basis functions) during an MD simulation (ML_MODE = train). However, at the start of the trajectory usually only very few local reference configurations are available. How many depends on the system size, symmetry and number of atoms (per type). This tag controls how many local reference configurations are required for each atom type before training of a force field is allowed. Setting a higher value may yield a more robust initial force field at the cost of more ab initio calculations during the first few MD steps.

If a training is scheduled in the current step, i.e., the STATUS line in the ML_LOGFILE shows learning or critical, but not enough atomic environments were collected, a text message will be emitted in the MSG line. Here is an example:

--------------------------------------------------------------------------------
STATUS                  1  critical  4      T      F         0         1
LCONF                   1 Ni         0         4
SPRSC                   1         1         1 Ni         4         1
MSG                     1      info Number of local reference configurations after sparsification below ML_MB_MIN, skipping training.
BEEF                    1   1.00000000E-06   3.46410162E-02   2.00000000E-02   2.00000000E-03   0.00000000E+00   0.00000000E+00
--------------------------------------------------------------------------------
STATUS                  2  critical  4      T      F         0         2
LCONF                   2 Ni         1         5
SPRSC                   2         2         2 Ni         5         2
MSG                     2      info Number of local reference configurations after sparsification below ML_MB_MIN, skipping training.
BEEF                    2   1.00000000E-06   3.46410162E-02   2.00000000E-02   2.00000000E-03   0.00000000E+00   0.00000000E+00
--------------------------------------------------------------------------------
STATUS                  3  critical  4      T      F         0         3
LCONF                   3 Ni         2         6
SPRSC                   3         3         3 Ni         6         3
REGR                    3    1    1   6.42271768E+01   1.36891529E+00   6.43938951E-14   3.76458019E+01
REGR                    3    1    2   7.43557292E+01   1.35099932E+00   5.48943502E-14   3.71479653E+01
REGR                    3    1    3   7.81437752E+01   1.34829112E+00   5.21286212E-14   3.70727138E+01
REGRF                   3    1    4   7.95268527E+01   1.34739777E+00   5.11880969E-14   3.70478685E+01    2.31498305E+15   3.30988608E+11
NDESC                   3        12 Ni       113
NDESC_SIC               3 Ni       113
STDAB                   3   1.41895828E-04   3.22944034E-02   9.66711031E-02
ERR                     3   1.45620371E-04   3.32055814E-02   9.84035846E-02
CFE                     3   0.00000000E+00   0.00000000E+00   0.00000000E+00
LASTE                   3   1.99936842E-04   5.14296619E-02   1.37823063E-01
BEE                     3   1.00000000E-06   3.46410162E-02   2.00000000E-02   2.00000000E-03   0.00000000E+00   0.00000000E+00
BEEF                    3   2.37926034E-05   2.21630107E-03   8.49321832E-04   2.00000000E-03   5.19041973E-02   3.66970792E-02
-------------------------------------------------------------------------------- 
...

Here, in the first two steps the ML_MB_MIN threshold prevents training of an MLFF, even if the Bayesian error estimation signals a critical step. The force field is then only generated in the third step, after the minimum number of three local reference configurations have been collected.

Related tags and articles

ML_LMLFF, ML_MODE, ML_MB, ML_MCONF_NEW, ML_MCONF, ML_LBASIS_DISCARD

Examples that use this tag