ML SIGW0: Difference between revisions

From VASP Wiki
(small typo corrected)
(Describe useful parameter range. Point out instability issues.)
Line 5: Line 5:
Description: This flag sets the precision parameter <math>s_{\mathrm{w}}</math> (see [[Machine learning force field: Theory#Bayesian linear regression|here]] for definition) for the fitting in the machine learning force field method.  
Description: This flag sets the precision parameter <math>s_{\mathrm{w}}</math> (see [[Machine learning force field: Theory#Bayesian linear regression|here]] for definition) for the fitting in the machine learning force field method.  
----
----
The default value for {{TAG|ML_MODE}}=''REFIT'' works reliably in most calculations.
The default value for {{TAG|ML_MODE}}=''REFIT'' works reliably in most calculations, however, sometimes it is necessary to increase the regularization parameter to avoid instabilities during finite temperature molecular dynamics simulations.


However, if the regularization needs to be controlled manually, like e.g. in the fitting via singular value decomposition ({{TAG|ML_MODE}}=''REFIT'' or {{TAG|ML_IALGO_LINREG}}=4), the best is to control the regularization via this parameter and keep the noise paramter <math>s_{\mathrm{v}}</math> (see {{TAG|ML_SIGV0}}) constant at 1.
If the regularization needs to be controlled manually, like e.g. in the fitting via singular value decomposition ({{TAG|ML_MODE}}=''REFIT'' or {{TAG|ML_IALGO_LINREG}}=4), the best is to control the regularization via this parameter and keep the noise parameter <math>s_{\mathrm{v}}</math> (see {{TAG|ML_SIGV0}}) constant at 1. Usefull values for {{TAG|ML_SIGW0}} are typically between 1E-2 and 1E-9. When the value is increased the regularization is increased. This often increases the training set error slightly but also reduces instabilities. If instabilities are observed during molecular dynamics simulations, one should therefore try to increase {{TAG|ML_SIGW0}} to say 1E-2, refit, and repeat the molecular dynamics simulations. 


For the theory of this regularization parameter see [[Machine learning force field: Theory#Regression|this section]].
More on the theory of this regularization parameter can be found in [[Machine learning force field: Theory#Regression|this section]].
== Related tags and articles ==
== Related tags and articles ==
{{TAG|ML_LMLFF}}, {{TAG|ML_MODE}}, {{TAG|ML_IREG}}, {{TAG|ML_SIGV0}}, {{TAG|ML_IALGO_LINREG}}
{{TAG|ML_LMLFF}}, {{TAG|ML_MODE}}, {{TAG|ML_IREG}}, {{TAG|ML_SIGV0}}, {{TAG|ML_IALGO_LINREG}}

Revision as of 14:31, 27 September 2024

ML_SIGW0 = [real]
Default: none 

Default: ML_SIGW0 = 1E-7 for ML_MODE = REFIT
= 1.0 else

Description: This flag sets the precision parameter (see here for definition) for the fitting in the machine learning force field method.


The default value for ML_MODE=REFIT works reliably in most calculations, however, sometimes it is necessary to increase the regularization parameter to avoid instabilities during finite temperature molecular dynamics simulations.

If the regularization needs to be controlled manually, like e.g. in the fitting via singular value decomposition (ML_MODE=REFIT or ML_IALGO_LINREG=4), the best is to control the regularization via this parameter and keep the noise parameter (see ML_SIGV0) constant at 1. Usefull values for ML_SIGW0 are typically between 1E-2 and 1E-9. When the value is increased the regularization is increased. This often increases the training set error slightly but also reduces instabilities. If instabilities are observed during molecular dynamics simulations, one should therefore try to increase ML_SIGW0 to say 1E-2, refit, and repeat the molecular dynamics simulations.

More on the theory of this regularization parameter can be found in this section.

Related tags and articles

ML_LMLFF, ML_MODE, ML_IREG, ML_SIGV0, ML_IALGO_LINREG

Examples that use this tag