ML EPS LOW: Difference between revisions
(Updated recommendation for value range) |
|||
| Line 5: | Line 5: | ||
---- | ---- | ||
This value sets the threshold for the eigenvalues that contribute to the leverage scoring used in the CUR algorithm for the rank compression ("sparsification") of the local configurations (for details see appendix E of reference {{cite|jinnouchi2:arx:2019}}). Small eigenvalues and those columns (local configurations) that are strongly connected | This value sets the threshold for the eigenvalues that contribute to the leverage scoring used in the CUR algorithm for the rank compression ("sparsification") of the local configurations (for details see appendix E of reference {{cite|jinnouchi2:arx:2019}}). Small eigenvalues and those columns (local configurations) that are strongly connected | ||
with these small eigenvalues are removed by the sparsification routines. The default value is fairly well balanced, and we do not recommend to | with these small eigenvalues are removed by the sparsification routines. The default value is fairly well balanced, and we do not recommend to increase the threshold to values | ||
larger than 1E-7. Also using smaller values than 1E-9 does not improve the MLFF if Bayesian regression is used | |||
(but it can be benficial for SVD). | |||
The description how to choose {{TAG|ML_EPS_LOW}} for accurate force fields is given [[Machine learning force field calculations: Basics#Accurate force fields|here]]. | The description how to choose {{TAG|ML_EPS_LOW}} for accurate force fields is given [[Machine learning force field calculations: Basics#Accurate force fields|here]]. | ||
Revision as of 12:36, 27 April 2022
ML_EPS_LOW = [real]
Default: ML_EPS_LOW = 1E-9 (vasp.6.3.1 default was 1E-10, see comments below)
Description: Threshold for the CUR algorithm used in the sparsification of local reference configurations within the machine learning force fields.
This value sets the threshold for the eigenvalues that contribute to the leverage scoring used in the CUR algorithm for the rank compression ("sparsification") of the local configurations (for details see appendix E of reference [1]). Small eigenvalues and those columns (local configurations) that are strongly connected with these small eigenvalues are removed by the sparsification routines. The default value is fairly well balanced, and we do not recommend to increase the threshold to values larger than 1E-7. Also using smaller values than 1E-9 does not improve the MLFF if Bayesian regression is used (but it can be benficial for SVD).
The description how to choose ML_EPS_LOW for accurate force fields is given here.
On the theory of the sparsification of local reference configurations see here.
Related tags and articles
ML_LMLFF, ML_MB, ML_EPS_REG, ML_IALGO_LINREG
References