ML ISTART: Difference between revisions
No edit summary |
No edit summary |
||
| Line 7: | Line 7: | ||
*{{TAG|ML_ISTART}} = 0: On-the-fly learning is enabled, starting from scratch. Force predictions from the machine learning force field are used to drive the MD simulation. However, if the error estimation performed in each time step indicates a high force error an ab initio calculation is performed instead and the collected energy, forces and stress are used to improve the machine learning force field. Setting {{TAG|ML_ISTART}} = 0 starts the machine learning force field from scratch. Hence, in the beginning of the MD run there is no force field available and ab initio calculation will happen frequently. | *{{TAG|ML_ISTART}} = 0: On-the-fly learning is enabled, starting from scratch. Force predictions from the machine learning force field are used to drive the MD simulation. However, if the error estimation performed in each time step indicates a high force error an ab initio calculation is performed instead and the collected energy, forces and stress are used to improve the machine learning force field. Setting {{TAG|ML_ISTART}} = 0 starts the machine learning force field from scratch. Hence, in the beginning of the MD run there is no force field available and ab initio calculation will happen frequently. | ||
*{{TAG|ML_ISTART}} = 1: Same as {{TAG|ML_ISTART}} = 0 but taking into account pre-existing ab initio data. Before the MD run starts the {{FILE|ML_AB}} file is read and the containing ab initio energies, forces and stresses are used to generate an initial force field. Then, the MD simulation is started with on-the-fly learning enabled. | *{{TAG|ML_ISTART}} = 1: Same as {{TAG|ML_ISTART}} = 0 but taking into account pre-existing ab initio data. Before the MD run starts the {{FILE|ML_AB}} file is read and the containing ab initio energies, forces and stresses are used to generate an initial force field. Then, the MD simulation is started with on-the-fly learning enabled. | ||
*{{TAG|ML_ISTART}} = 2: Prediction only. In this mode the previously trained machine learning force field is read from the {{FILE|ML_FF}} file. The MD simulation is driven with predictions from the force field only, no ab initio calculations are performed and no learning is executed. The | *{{TAG|ML_ISTART}} = 2: Prediction only. In this mode the previously trained machine learning force field is read from the {{FILE|ML_FF}} file. The MD simulation is driven with predictions from the force field only, no ab initio calculations are performed and no learning is executed. | ||
*{{TAG|ML_ISTART}} = 3: Learning from given ab initio data only, no MD time steps. In this operation mode a new machine learning force field is generated from ab initio data provided in the {{FILE|ML_AB}} file. The structures are read in and processed one by one as if harvested via an MD simulation. Hence, if an {{FILE|ML_AB}} file from a previous on-the-fly training is supplied this will recreate the same force field, i.e. output in the {{FILE|ML_FF}} file, without repetition of the ab initio calculations. This is particularly useful for testing different machine learning force field parameters. | |||
{{NB|warning|This feature is experimental!|:}} | {{NB|mind|Consider a force field generated with on-the-fly learning and a specific set of parameters (e.g. cutoff radii, number of radial basis functions, etc.), called settings "A". This will result in a force field in {{FILE|ML_FF}} and an ab initio data set of structures in {{FILE|ML_AB}}. Using {{TAG|ML_ISTART}} {{=}} 3 allows to easily test out different parameters, i.e. settings "B", given only the existing ab initio data from the original on-the-fly run. No new ab initio calculations are performed and in the end another force field is created. However, this new force field will usually '''not''' be identical to a third one generated directly with on-the-fly learning using settings "B". This is because the trajectories will quickly diverge when settings "A" and "B" result in different force predictions. Therefore, the ab initio data sets collected from "A" and "B" on-the-fly runs will not match.|:}} | ||
{{NB|warning|This feature is experimental! Problems may arise if the {{FILE|ML_AB}} contains structures with mixed number of elements and atom numbers.|:}} | |||
== Related Tags and Sections == | == Related Tags and Sections == | ||
Revision as of 00:21, 8 November 2021
ML_ISTART = [integer]
Default: ML_ISTART = 0
Description: This tag selects the mode of operation (e.g. start from scratch, prediction-only,...) of the machine learning force fields method.
If the machine learning force fields method is enabled via ML_LMLFF = .TRUE., this tag further specifies the mode of operation when VASP is run. The following cases can be selected:
- ML_ISTART = 0: On-the-fly learning is enabled, starting from scratch. Force predictions from the machine learning force field are used to drive the MD simulation. However, if the error estimation performed in each time step indicates a high force error an ab initio calculation is performed instead and the collected energy, forces and stress are used to improve the machine learning force field. Setting ML_ISTART = 0 starts the machine learning force field from scratch. Hence, in the beginning of the MD run there is no force field available and ab initio calculation will happen frequently.
- ML_ISTART = 1: Same as ML_ISTART = 0 but taking into account pre-existing ab initio data. Before the MD run starts the ML_AB file is read and the containing ab initio energies, forces and stresses are used to generate an initial force field. Then, the MD simulation is started with on-the-fly learning enabled.
- ML_ISTART = 2: Prediction only. In this mode the previously trained machine learning force field is read from the ML_FF file. The MD simulation is driven with predictions from the force field only, no ab initio calculations are performed and no learning is executed.
- ML_ISTART = 3: Learning from given ab initio data only, no MD time steps. In this operation mode a new machine learning force field is generated from ab initio data provided in the ML_AB file. The structures are read in and processed one by one as if harvested via an MD simulation. Hence, if an ML_AB file from a previous on-the-fly training is supplied this will recreate the same force field, i.e. output in the ML_FF file, without repetition of the ab initio calculations. This is particularly useful for testing different machine learning force field parameters.
Mind: Consider a force field generated with on-the-fly learning and a specific set of parameters (e.g. cutoff radii, number of radial basis functions, etc.), called settings "A". This will result in a force field in ML_FF and an ab initio data set of structures in ML_AB. Using ML_ISTART = 3 allows to easily test out different parameters, i.e. settings "B", given only the existing ab initio data from the original on-the-fly run. No new ab initio calculations are performed and in the end another force field is created. However, this new force field will usually not be identical to a third one generated directly with on-the-fly learning using settings "B". This is because the trajectories will quickly diverge when settings "A" and "B" result in different force predictions. Therefore, the ab initio data sets collected from "A" and "B" on-the-fly runs will not match.
Warning: This feature is experimental! Problems may arise if the ML_AB contains structures with mixed number of elements and atom numbers.
Related Tags and Sections
ML_LMLFF, ML_W1, ML_WTOTEN, ML_WTIFOR, ML_WTSIF, ML_IALGO_LINREG