ML SCLC CTIFOR: Difference between revisions

Latest revision as of 12:33, 12 June 2024

ML_SCLC_CTIFOR = [real]
Default: ML_SCLC_CTIFOR = 0.6

Description: Sets fraction by which the error threshold for the maximum forces is lowered in the selection of local reference calculations.

ML_CTIFOR determines whether a first-principles calculation is performed when training an MLFF. Whenever a first-principles calculation is performed, additional functions are added to the sparse representation of the kernel (local-reference configurations). ML_SCLC_CTIFOR determines how many local-reference configurations are added to the sparse representation of the kernel. Specifically, the local environment of those atoms with an estimated error larger than ML_SCLC_CTIFOR * ML_CTIFOR are added as candidates for the sparse representational of the kernel. Note that changing ML_SCLC_CTIFOR does not change the decision of whether a first-principles calculation is carried out or not, since this decision is entirely based on ML_CTIFOR.

The default value of 0.6 is often a reasonably good compromise. If the value is decreased, obviously more functions are used for the sparse representation of the kernel. This always improves the initial learning efficiency but might slow the force-field calculations. So ML_SCLC_CTIFOR compromises either learning efficiency or the speed of the evaluation of the MLFF. For polymers and liquids, we found that decreasing ML_SCLC_CTIFOR to values around 0.4 (or even smaller) can significantly improve learning efficiency.

@@ Line 1: / Line 1: @@
 {{DISPLAYTITLE:ML_SCLC_CTIFOR}}
-{{TAGDEF|ML_IREG|[real]|0.8}}
+{{TAGDEF|ML_SCLC_CTIFOR|[real]|0.6}}
-Description: Sets fraction by which the Bayesian threshold for the maximum forces is lowered in the selection of local reference calculations.
+Description: Sets fraction by which the error threshold for the maximum forces is lowered in the selection of local reference calculations.
 ----
-At every sampling step each atom with it's environment is chosen as a local reaference configuration is chosen if it's predicted error in the force is larger than the Bayesian threshold for the maximum forces ({{TAG|ML_CTIFOR}}). {{TAG|ML_SCLC_CTIFOR}} lowers the threshold by multiplying it to {{TAG|ML_CTIFOR}}, but it doesn't effect the decision whether a sampling step (ab-initio calculation) is carried out or not, since the decision for sampling is done previously in a separate step.
+{{TAG|ML_CTIFOR}} determines whether a first-principles calculation is performed when training an MLFF.
+Whenever a first-principles calculation is performed, additional functions are added to the sparse representation of the kernel (local-reference configurations). {{TAG | ML_SCLC_CTIFOR}} determines how many local-reference configurations are added to the sparse representation of the kernel.
+Specifically, the local environment of those atoms with an estimated error larger than {{TAG | ML_SCLC_CTIFOR}} * {{TAG|ML_CTIFOR}} are
+added as candidates for the sparse representational of the kernel. Note that changing {{TAG | ML_SCLC_CTIFOR}} does not change the decision of whether a first-principles calculation is carried out or not, since this decision is entirely based on {{TAG|ML_CTIFOR}}.
-The default value of 0.8 was set by us empirically in dependence of default values for other input parameters.
+The default value of 0.6 is often a reasonably good compromise. If the value is decreased, obviously more functions are used for the sparse representation of the kernel. This always improves the initial learning efficiency but might slow the force-field calculations.
+So {{TAG | ML_SCLC_CTIFOR}} compromises either learning efficiency or the speed of the evaluation of the MLFF.
-If this value is decreased, often the initial learning efficiency is much improved. This applies in particular to liquid and polymeric system, with a large configurational space. Good values are around 0.5 for such systems.
+For polymers and liquids, we found that decreasing {{TAG | ML_SCLC_CTIFOR}} to values around  0.4 (or even smaller) can significantly improve learning efficiency.
 == Related tags and articles ==