ML CX: Difference between revisions
mNo edit summary |
No edit summary |
||
| Line 6: | Line 6: | ||
The use of this tag in combination with the learning algorithms is described here: [[Machine learning force field calculations: Basics#Threshold for error of forces|here]]. | The use of this tag in combination with the learning algorithms is described here: [[Machine learning force field calculations: Basics#Threshold for error of forces|here]]. | ||
If {{TAG|ML_ICRITERIA}}>0, {{TAG|ML_CTIFOR}} is set to the average of the | If {{TAG|ML_ICRITERIA}}>0, {{TAG|ML_CTIFOR}} is set to the average of the errors of the forces stored in a history. Note that {{TAG|ML_ICRITERIA}}=1 and {{TAG|ML_ICRITERIA}}=2, average over different data. In the first case the average is performed over errors after updates of the force fields, and in the second case over all recent error estimates (see {{TAG|ML_ICRITERIA}}). In both cases, if {{TAG|ML_CTIFOR}} is updated, it is set to | ||
{{TAG|ML_CTIFOR}} = (average of the stored | {{TAG|ML_CTIFOR}} = (average of the stored errors in the history) *(1.0 + {{TAG|ML_CX}}). | ||
Obviously setting {{TAG|ML_CX}} to a positive value will result in fewer first principles calculations and fewer updates of the MLFF, whereas negative values result in more frequent first principles calculations (as well as updates of the MLFF). Typical values of {{TAG|ML_CX}} are between -0.2 and 0.0 for {{TAG|ML_ICRITERIA}}=1, and 0.0 and 0.3 for {{TAG|ML_ICRITERIA}}=2 (a good starting value is 0.2 for {{TAG|ML_ICRITERIA}}=2). For training runs using heating, the default usually results in very well balanced machine learned force fields. When the training is performed at a fixed temperature, it is often desirable to decrease to {{TAG|ML_CX}}=-0.1, in order to increase the number of first principle calculations and thus the size of the training set (the default can result in too few training data). | Obviously setting {{TAG|ML_CX}} to a positive value will result in fewer first principles calculations and fewer updates of the MLFF, whereas negative values result in more frequent first principles calculations (as well as updates of the MLFF). Typical values of {{TAG|ML_CX}} are between -0.2 and 0.0 for {{TAG|ML_ICRITERIA}}=1, and 0.0 and 0.3 for {{TAG|ML_ICRITERIA}}=2 (a good starting value is 0.2 for {{TAG|ML_ICRITERIA}}=2). For training runs using heating, the default usually results in very well balanced machine learned force fields. When the training is performed at a fixed temperature, it is often desirable to decrease to {{TAG|ML_CX}}=-0.1, in order to increase the number of first principle calculations and thus the size of the training set (the default can result in too few training data). | ||
Latest revision as of 12:26, 12 June 2024
ML_CX = [real]
Default: ML_CX = 0.0
Description: The parameter determines to which value the threshold (ML_CTIFOR) is updated within the machine learning force field methods.
The use of this tag in combination with the learning algorithms is described here: here.
If ML_ICRITERIA>0, ML_CTIFOR is set to the average of the errors of the forces stored in a history. Note that ML_ICRITERIA=1 and ML_ICRITERIA=2, average over different data. In the first case the average is performed over errors after updates of the force fields, and in the second case over all recent error estimates (see ML_ICRITERIA). In both cases, if ML_CTIFOR is updated, it is set to
ML_CTIFOR = (average of the stored errors in the history) *(1.0 + ML_CX).
Obviously setting ML_CX to a positive value will result in fewer first principles calculations and fewer updates of the MLFF, whereas negative values result in more frequent first principles calculations (as well as updates of the MLFF). Typical values of ML_CX are between -0.2 and 0.0 for ML_ICRITERIA=1, and 0.0 and 0.3 for ML_ICRITERIA=2 (a good starting value is 0.2 for ML_ICRITERIA=2). For training runs using heating, the default usually results in very well balanced machine learned force fields. When the training is performed at a fixed temperature, it is often desirable to decrease to ML_CX=-0.1, in order to increase the number of first principle calculations and thus the size of the training set (the default can result in too few training data).
The number of entries in the history are controlled by ML_MHIS for ML_ICRITERIA=1, and it is currently fixed to 400 for ML_ICRITERIA=2 (in future releases 50 x ML_MHIS).
Related tags and articles
ML_LMLFF, ML_ICRITERIA, ML_CTIFOR, ML_MHIS, ML_CSIG, ML_CSLOPE