ML SIGW0: Difference between revisions

Latest revision as of 14:36, 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 $s_{\mathrm{w}}$ (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.

Suppose 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 $s_{\mathrm{v}}$ (see ML_SIGV0) constant at 1. Useful values for ML_SIGW0 are typically between 1E-1 and 1E-9. When the value is increased the regularization is increased. This often slightly increases the training set error but also reduces potential instabilities. Hence, if instabilities are observed during molecular-dynamics simulations, one should try to increase ML_SIGW0 to say 1E-3, then refit, and repeat the molecular-dynamics simulations.

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

@@ Line 3: / Line 3: @@
 {{DEF|ML_SIGW0|1E-7|for {{TAG|ML_MODE}} {{=}} REFIT|1.0|else}}
-Description: This flag sets the precision parameter <math>s_{\mathrm{w}}</math> 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 works empirically very good 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.
+Suppose 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. Useful values for {{TAG|ML_SIGW0}} are typically between 1E-1 and 1E-9. When the value is increased the regularization is increased. This often slightly increases the training set error but also reduces potential instabilities. Hence, if instabilities are observed during molecular-dynamics simulations, one should try to increase {{TAG|ML_SIGW0}} to say 1E-3, then 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 ==
 {{TAG|ML_LMLFF}}, {{TAG|ML_MODE}}, {{TAG|ML_IREG}}, {{TAG|ML_SIGV0}}, {{TAG|ML_IALGO_LINREG}}