ML MHIS: Difference between revisions

Latest revision as of 12:28, 12 June 2024

ML_MHIS = [integer]
Default: ML_MHIS = 10

Description: This tag sets the number of estimated errors stored in memory to determine the threshold for error estimation in the machine learning force field method for ML_ICRITERIA=1. For ML_ICRITERIA=2, the history length is 50 x ML_MHIS (or hard coded to 400).

The use of this tag in combination with the learning algorithms is described here: here.

ML_ICRITERIA=1: The ML code stores ML_MHIS errors from previous training steps: immediately after a re-training of the ML-FF, the estimated errors of the forces are reevaluated for the current structure (that was also just added as training structure). The average and the maximum error of the forces is stored in the history. After ML_MHIS updates of the force field, the threshold ML_CTIFOR is updated the first time and is then updated after every further update of the ML-FF. We recommend to read the section ML_ICRITERIA for further details.

ML_ICRITERIA=2: Averaging is performed over 50 x ML_MHIS error predictions. Every MD step is considered in the averaging (as opposed to above, where only structures after re-training are considered).

@@ Line 2: / Line 2: @@
 {{TAGDEF|ML_MHIS|[integer]|10}}
-Description: This tag sets the number of estimated errors stored in memory to determine the threshold for the Bayesian error in the machine learning force field method for {{TAG|ML_ICRITERIA}}=1. For {{TAG|ML_ICRITERIA}}=2, the history length is 50 x {{TAG| ML_MHIS}} (or hard coded to 400).
+Description: This tag sets the number of estimated errors stored in memory to determine the threshold for error estimation in the machine learning force field method for {{TAG|ML_ICRITERIA}}=1. For {{TAG|ML_ICRITERIA}}=2, the history length is 50 x {{TAG| ML_MHIS}} (or hard coded to 400).
 ----
 The use of this tag in combination with the learning algorithms is described here: [[Machine_learning_force_field:_Theory#Threshold_for_error_of_forces|here]].
-{{TAG|ML_ICRITERIA}}=1: The ML code stores {{TAG|ML_MHIS}} Bayesian errors from previous training steps: immediately after a re-training of the ML-FF, the Bayesian errors of the forces are reevaluated for the current structure (that was also just added as training structure). The average  and the maximum error of the forces is stored in the history. After  {{TAG | ML_MHIS}} updates of the force field, the threshold   {{TAG|ML_CTIFOR}} is updated the first time and is then updated after every further update of the ML-FF. We recommend to read the section {{TAG|ML_ICRITERIA}} for further details.
+{{TAG|ML_ICRITERIA}}=1: The ML code stores {{TAG|ML_MHIS}} errors from previous training steps: immediately after a re-training of the ML-FF, the estimated errors of the forces are reevaluated for the current structure (that was also just added as training structure). The average  and the maximum error of the forces is stored in the history. After  {{TAG | ML_MHIS}} updates of the force field, the threshold   {{TAG|ML_CTIFOR}} is updated the first time and is then updated after every further update of the ML-FF. We recommend to read the section {{TAG|ML_ICRITERIA}} for further details.
-{{TAG|ML_ICRITERIA}}=2: Averaging is performed over 50 x {{TAG| ML_MHIS}} Bayesian error predictions. Every MD step is considered in the averaging (as opposed to above, where only structures after re-training are considered).
+{{TAG|ML_ICRITERIA}}=2: Averaging is performed over 50 x {{TAG| ML_MHIS}} error predictions. Every MD step is considered in the averaging (as opposed to above, where only structures after re-training are considered).
 == Related tags and articles ==