Input

Simple Active Learning reads all options from an input file, described here. You can also set up this input file in Python.

Table 1 AMS-like Molecular Dynamics input options

Block

Required?

Comment

System/LoadSystem

Yes

identical to AMS Driver System/LoadSystem

Task

No

must be set to MolecularDynamics

MolecularDynamics

Yes

identical to AMS Driver MolecularDynamics

Constraints

No

identical to AMS Driver Constraints

RNGSeed

No

Random number seed(s) for MD simulations

Engine

Yes

reference engine settings, identical to normal AMS calculations

MachineLearning

Yes

identical to ParAMS MachineLearning settings

ParallelLevels

No

identical to ParAMS ParallelLevels settings

ActiveLearning

Yes

described on this page

The engine settings for the MD simulations are determined from the MachineLearning input. For example, if you train an M3GNet model, this means that you will automatically run M3GNet also during the MD simulation.

This section only describes the ActiveLearning input block, which controls

  • How to generate/load initial reference data

  • When to perform reference calculations

  • Criteria for deciding whether to retrain the model or continue the MD simulation

  • How much output to save

  • Whether to retrain the model and/or rerun the simulation after the active learning loop has finished

See also

Overview

ActiveLearning
   AtEnd
      RerunSimulation Yes/No
      RetrainModel Yes/No
   End
   InitialReferenceData
      Generate
         M3GNetShortMD
            Enabled Yes/No
         End
         ReferenceMD
            Enabled Yes/No
         End
      End
      Load
         Directory string
         FromPreviousModel Yes/No
      End
   End
   JobPrefix string
   MaxAttemptsPerStep integer
   MaxReferenceCalculationsPerAttempt integer
   ReasonableSimulationCriteria
      Distance
         Enabled Yes/No
         MinValue float
      End
      EnergyUncertainty
         Enabled Yes/No
         MaxValue float
         Normalization float
      End
      GradientsUncertainty
         Enabled Yes/No
         MaxValue float
      End
      Temperature
         Enabled Yes/No
         MaxValue float
      End
   End
   Save
      ReferenceCalculations [None | All]
      ReferenceData [Latest | All]
      TrainingDirectories [Latest | All]
      Trajectories [Latest | All]
   End
   Steps
      Geometric
         NumSteps integer
         Start integer
      End
      Linear
         Start integer
         StepSize integer
      End
      List integer_list
      Type [Geometric | List | Linear]
   End
   SuccessCriteria
      Energy
         Enabled Yes/No
         Normalization float
         Relative float
         Total float
      End
      Forces
         Enabled Yes/No
         MaxDeviationForZeroForce float
         MaxMAE float
         MinR2 float
      End
   End
End

Initial reference data

Before the main active learning loop starts, there must be some training data.

The initial training data can be loaded from disk and/or automatically generated. If no data is loaded and no generation option is explicitly enabled, then the ReferenceMD option described below will be automatically enabled to ensure that there is some data for the initial model training.

ActiveLearning
   InitialReferenceData
      Generate
         M3GNetShortMD
            Enabled Yes/No
         End
         ReferenceMD
            Enabled Yes/No
         End
      End
      Load
         Directory string
         FromPreviousModel Yes/No
      End
   End
End

Generate initial reference data

The M3GNetShortMD option (recommended) follows a short pre-programmed MD simulation using the universal M3GNet-UP-2022 potential. This gives some structural variation in the initial training data. It generates structures as follows:

  • 300 MD steps with timestep 0.5 fs, temperature = 500 K

  • If the system is 3d-periodic then linearly scale the density from 92% to 108% of the original density

  • 5 frames are recalculated with the reference engine and added to the training/validation sets

The ReferenceMD option (default if nothing else is specified)

  • Runs 3 MD steps (saving every frame) using the exact MolecularDynamics settings specified in the input

  • Adds those frames to the training/validation sets

Load initial reference data

If you already have some reference data, for example if you have

  • previously run Simple Active Learning, or

  • manually created the data by importing into ParAMS and saving,

then you can load it in Simple Active Learning, so that the old data is combined with the new data generated during the workflow.

If you specify the ActiveLearning%InitialReferenceData%Load%Directory option, then the initial reference data will be taken from that directory.

Otherwise, if you’re loading a previously trained model using MachineLearning%LoadModel, and if you enable ActiveLearning%InitialReferenceData%Load%FromPreviousModel, then both the parameters and the training and validation data will be loaded.

Initial reference data input

ActiveLearning
Type:

Block

Description:

Settings for Active Learning

InitialReferenceData
Type:

Block

Description:

Options for loading reference data.

Generate
Type:

Block

Description:

How to generate initial reference data from the initial structure. Can also be combined with the Load block. The purpose of these options is to get some initial reference structures/data around the current structure that can be used for Step 1 of the active learning loop. The ReferenceMD option will be automatically enabled if no data is otherwise loaded or generated.

M3GNetShortMD
Type:

Block

Description:

Structure sampler using M3GNet-UP-2022

Enabled
Type:

Bool

Default value:

No

GUI name:

M3GNet-UP short MD:

Description:

Run 300 steps with M3GNet-UP-2022 at T=600 K. If the system is 3D-periodic the density will be scanned around the initial value. Extract 5 frames and run reference calculations on those.

ReferenceMD
Type:

Block

Description:

Run NSteps of the MD simulation using the reference engine.

Enabled
Type:

Bool

Default value:

No

GUI name:

Reference MD:

Description:

Run 3 steps with the reference engine and add those 3 frames to the training and validation sets. If no other reference data is loaded or generated, this option will automatically be enabled.

Load
Type:

Block

Description:

How to load initial reference data from other sources. Can also be combined with the Generate block

Directory
Type:

String

Default value:

Description:

Directory containing initial reference data. It can be * a ParAMS input directory or a stepX_attemptY_reference_data directory containing the files job_collection.yaml, training_set.yaml, and validation_set.yaml. * a ParAMS results directory. If a directory is specified here it will be used instead of the data from a previously loaded model.

FromPreviousModel
Type:

Bool

Default value:

Yes

Description:

If MachineLearning%LoadModel is set, reuse reference data from that ParAMS run. If MachineLearning%LoadModel is not set, or if Directory is specified, then this input option is ignored.

When to run reference calculations (step sequence type)

In the Simple Active Learning workflow, the MD simulation is divided into a sequence of active learning (AL) steps.

ActiveLearning
   Steps
      Geometric
         NumSteps integer
         Start integer
      End
      Linear
         Start integer
         StepSize integer
      End
      List integer_list
      Type [Geometric | List | Linear]
   End
   MaxAttemptsPerStep integer
   MaxReferenceCalculationsPerAttempt integer
   JobPrefix string
End

Step Type Geometric (default)

Example:

  • You set up the MD simulation with NMD = 10000 steps with a time step of 0.5 fs, giving a total simulation length of 10000*0.5 = 5000 fs = 5 ps.

  • You set up the ActiveLearning with Steps%Type = Geometric with Start set to 10 (MD frames) and NumSteps set to 5, and MaxAttemptsPerStep set to 8

For example using the following input:

MolecularDynamics
    NSteps 10000
    TimeStep 0.5
    # ... other MD options
End

ActiveLearning
    Steps
        Type Geometric   # default
        Geometric
            Start 10     # default
            NumSteps 5
        End
    End
    MaxAttemptsPerStep 8
    MaxReferenceCalculationsPerAttempt 4
    # ... other ActiveLearning options
End

This will divide the 10000 MD steps into 5 AL steps, where the first AL step contains 10 MD steps, and each subsequent AL step contains progressively more MD steps (following a Geometric progression):

The ACTIVE LEARNING loop will contain 5 steps, using the following scheme:
   Active Learning Step   1:       10 MD Steps (cumulative:       10)
   Active Learning Step   2:       46 MD Steps (cumulative:       56)
   Active Learning Step   3:      260 MD Steps (cumulative:      316)
   Active Learning Step   4:     1462 MD Steps (cumulative:     1778)
   Active Learning Step   5:     8222 MD Steps (cumulative:    10000)
Total number of MD Steps: 10000
Max attempts per active learning step: 8

The progression is geometric because 56/10 ≈ 316/56 ≈ 1778/316 ≈ 10000/1778 ≈ 5.6.

The above scheme means that the active learning loop will be executed as follows:

  1. step1_attempt1_simulation: Run 10 MD steps using the initially trained model

  2. step1_attempt1_ref_calc1: Run reference calculation on final frame

  3. Evaluate the Success criteria:

  • If no success: run up to 3 more reference calculations, retrain the model, and loop back to the beginning of the step ↰: rerun AL step 1 (the first 10 MD steps) as step1_attempt2_simulation using the new parameters, run reference calculation on final frame, evaluate the success criteria, …

  • If success or if the number of attempts > 8: continue to AL step 2

  1. step2_attempt1_simulation: Run 46 MD steps starting from the final frame of AL step 1, for a total (cumulative) length of 56 MD steps

  2. step2_attempt1_ref_calc1: Run reference calculation on final frame

  3. Evaluate the Success criteria:

  • If no success: run up to 3 more reference calculations, retrain the model, and loop back to the beginning of the step ↰: rerun AL step 2 (the 46 MD steps) as step2_attempt2_simulation using the new parameters, run reference calculation on final frame, evaluate the success criteria, …

  • If success or if the number of attempts > 8: continue to AL step 3

  1. step3_attempt1_simulation: Run 260 MD steps starting from the final frame of AL step 2, for a total (cumulative) length of 315 MD steps

  2. Etcetera….

Step Type Linear

The steps can also follow a linear progression.

This is especially useful if you run non-equilibrium MD where you linearly apply some restraint, for example if you use a ReactionBoost RMSDRestraint following the TargetCoordinate, or apply a linear lattice deformation.

Instead of providing the number of steps, you provide the start step and the step size:

MolecularDynamics
    NSteps 10000
    # other MD options...
End

ActiveLearning
    Steps
        Type Linear
        Linear
            Start 100
            StepSize 2000
        End
    End
End

Active Learning Step   1:      100 MD Steps (cumulative:      100)
Active Learning Step   2:     2000 MD Steps (cumulative:     2100)
Active Learning Step   3:     2000 MD Steps (cumulative:     4100)
Active Learning Step   4:     2000 MD Steps (cumulative:     6100)
Active Learning Step   5:     2000 MD Steps (cumulative:     8100)
Active Learning Step   6:     1900 MD Steps (cumulative:    10000)

Step Type List

You can also list the (cumulative) number of MD steps per active learning step explicitly. The final MD step is always considered to be the end of an active learning step and does not need to be specified.

MolecularDynamics
    NSteps 10000
    # other MD options...
End

ActiveLearning
    Steps
        Type List
        List 100 3333 4567 7777
    End
End

Active Learning Step   1:      100 MD Steps (cumulative:      100)
Active Learning Step   2:     3233 MD Steps (cumulative:     3333)
Active Learning Step   3:     1234 MD Steps (cumulative:     4567)
Active Learning Step   4:     3210 MD Steps (cumulative:     7777)
Active Learning Step   5:     2223 MD Steps (cumulative:    10000)

Steps input

ActiveLearning
   Steps
      Geometric
         NumSteps integer
         Start integer
      End
      Linear
         Start integer
         StepSize integer
      End
      List integer_list
      Type [Geometric | List | Linear]
   End
   MaxAttemptsPerStep integer
   MaxReferenceCalculationsPerAttempt integer
   JobPrefix string
End
ActiveLearning
Type:

Block

Description:

Settings for Active Learning

Steps
Type:

Block

Description:

Settings to determine the number of MD steps per active learning step.

Geometric
Type:

Block

Description:

Options for geometric.

NumSteps
Type:

Integer

Default value:

10

Description:

The number of active learning steps to perform. The MD simulation will be split into this number of active learning steps. The active learning steps will progressively contain more and more MD steps.

Start
Type:

Integer

Default value:

10

Description:

The length of the first step (in MD time steps).

Linear
Type:

Block

Description:

Options for linear.

Start
Type:

Integer

Default value:

10

Description:

The length of the first step (in MD time steps).

StepSize
Type:

Integer

Default value:

1000

Description:

The length of every subsequent active learning step (in MD time steps).

List
Type:

Integer List

Description:

List of MD frame indices, for example 10 50 200 1000 10000 100000. Only indices smaller than MolecularDynamics%NSteps are considered. Note: the final frame of the MD simulation is always considered to be the end of a step and does not need to be specified here.

Type
Type:

Multiple Choice

Default value:

Geometric

Options:

[Geometric, List, Linear]

GUI name:

Step sequence type:

Description:

How to determine the number of MD steps per active learning step.

MaxAttemptsPerStep
Type:

Integer

Default value:

15

Description:

Maximum number of attempts per active learning step. If this number is exceeded, the active learning will continue to the next step even if the potential is not accurate enough according to the criteria. If the default value is exceeded, it probably means that the criteria are too strict.

MaxReferenceCalculationsPerAttempt
Type:

Integer

Default value:

4

GUI name:

Max ref calcs per attempt:

Description:

Maximum number of reference calculations per attempt. For successful attempts, only a single reference calculation is performed. For very short active learning steps, fewer calculations are done than the number specified.

JobPrefix
Type:

String

Default value:

Description:

Jobs added to the job collection will receive this prefix. Example: set to water_ to get jobs like water_step1_attempt1_frame001. If the prefix does not end with an underscore _, one will be automatically added.

Success criteria

At the end of an active learning step, a reference calculation (stepX_attemptY_ref_calc1) is performed on the last frame of the MD simulation.

The results (energy and forces) from this reference calculation are compared to the results of the trained ML potential.

Only if the agreement is accurate enough, such that all success criteria are fulfilled, will the Active Learning workflow continue to the next Active Learning Step.

Energy: total and relative

Enable the energy success checker with ActiveLearning%SuccessCriteria%Energy%Enabled.

Energies can optionally be normalized by some number before making the comparison, by specifying the ActiveLearning%SuccessCriteria%Energy%Normalization input option.

By default energies are normalized by the number of atoms. This is suitable for reasonably homogeneous systems and means that the same criteria can be used for any number of atoms.

You may consider changing the Normalization if your system is very inhomogeneous, for example if you’re looking at single atom diffusing in a large bulk crystal.

Total energy

The ActiveLearning%SuccessCriteria%Energy%Total compares the ML-predicted energy Epred directly to the reference energy Eref:

  • ΔE = Epred - Eref

  • Success if |ΔE|/Normalization < ActiveLearning%SuccessCriteria%Energy%Total

Relative energy

Compare the difference between calculated relative reference energies and relative predicted energies.

This success criterion is not invoked for step1_attempt1 but for all subsequent steps and attempts.

  • ΔEref = Erefcurrent - Erefprevious

  • ΔEpred = Epredcurrent - Epredprevious

  • ΔΔE = ΔEpred - ΔEref

  • Success if |ΔΔE|/Normalization < ActiveLearning%SuccessCriteria%Energy%Relative

Forces (gradients)

Enable the forces success criterion with ActiveLearning%SuccessCriteria%Forces%Enabled.

The predicted forces are compared to the reference forces in three ways:

  • Mean absolute error (MAE) in eV/angstrom, MaxMAE

  • R² in the correlation plot between reference and predicted values, MinR2

  • Maximum deviation, MaxDeviationForZeroForce

For structures with large components, it is usually not so important the the forces are predicted very accurately, as they represent unstable structures that are unlikely to appear in an MD simulation. For large force components, one can accept a larger error (deviation) between the reference and predicted values.

For this reason, the maximum deviation criterion depends on the magnitude of the reference force. The maximum allowed deviation between predicted and reference force components is determined by the following equation:

\[y(x) = y_0 + \frac{L}{1+\exp(-k(|x|-x_0))} - \frac{L}{1+\exp(-k(-x_0))}\]

where \(y\) is the threshold, \(x\) is the reference force, \(y_0\) is MaxDeviationForZeroForce, \(L\) = 3, \(x_0\) = 7, and \(k\) = 0.5.

There is no theoretical basis for this equation other than that it in practice seems to give reasonable thresholds.

This gives the following calculated threshold vs. reference force for a few different values of MaxDeviationForZeroForce:

/scm-uploads/doc.trunk/Workflows/_images/check_gradients_threshold_vs_reference_force.png

Success criteria input

ActiveLearning
   SuccessCriteria
      Energy
         Enabled Yes/No
         Normalization float
         Relative float
         Total float
      End
      Forces
         Enabled Yes/No
         MaxDeviationForZeroForce float
         MaxMAE float
         MinR2 float
      End
   End
End
ActiveLearning
Type:

Block

Description:

Settings for Active Learning

SuccessCriteria
Type:

Block

Description:

Criteria for determining whether an active learning step was successful. These criteria compare one or more reference calculations to the predictions. If any of the criteria are exceeded, the active learning loop will reparametrize the model and repeat the step.

Energy
Type:

Block

Description:

Conditions to decide whether the calculated energy is are accurate enough with respect to reference energies.

Enabled
Type:

Bool

Default value:

Yes

Description:

Enable energy checking during the active learning.

Normalization
Type:

Float

Description:

Normalize (divide) energies by this number before comparing to the specified thresholds. If not specified, it will become the number of atoms.

Relative
Type:

Float

Default value:

0.005

Unit:

eV

GUI name:

Relative energy:

Description:

|ΔΔE|/Normalization: Maximum allowed difference between the calculated relative reference energies and relative predicted energies. The relative energies are calculated for the current structure with respect to the structure in the previous reference calculation. ΔE_ref = E_ref(current) - E_ref(previous). ΔE_pred = E_pred(current) - E_pred(previous). |ΔΔE| = |ΔE_pred - ΔE_ref|

Total
Type:

Float

Default value:

0.2

Unit:

eV

GUI name:

Total energy:

Description:

|ΔE|/Normalization: Maximum allowed total energy difference between the reference and predicted energy. This criterion is mostly useful when restarting a workflow from a previously trained model but on a new stoichiometry / system, for which the total energy prediction may be very far from the target. The default value is quite large so it is normally not exceeded. |∆E| = |E_pred - E_ref|

Forces
Type:

Block

Description:

Conditions to decide whether calculated forces are accurate enough with respect to reference forces.

Enabled
Type:

Bool

Default value:

Yes

Description:

Enable checking the forces during the active learning.

MaxDeviationForZeroForce
Type:

Float

Default value:

0.5

Unit:

eV/angstrom

Description:

The maximum allowed deviation between a calculated force component and the corresponding reference force component. For larger reference forces, the allowed deviation will also be larger (see the documentation). If any deviation is larger than the (magnitude-dependent) threshold, the active learning step will be repeated after a reparametrization.

MaxMAE
Type:

Float

Default value:

0.3

Unit:

eV/angstrom

GUI name:

Max MAE:

Description:

Maximum allowed mean absolute error when comparing reference and predicted forces for a single frame at the end of an active learning step. If the obtained MAE is larger than this threshold, the active learning step will be repeated after a reparametrization.

MinR2
Type:

Float

Default value:

0.2

GUI name:

Min R²:

Description:

Minimum allowed value for R^2 when comparing reference and predicted forces for a single frame at the end of an active learning step. If the obtained R^2 is smaller than this threshold, the active learning step will be repeated after a reparametrization. Note that if you have very small forces (for example by running the active learning at a very low temperature or starting from a geometry-optimized structure), then you should decrease the MinR2 since it is difficult for the ML model predict very small forces accurately.

Reasonable simulation criteria (uncertainties, temperature, …)

When running MD simulations with ML potentials, it may happen that the simulation explores configurational space where the ML potential was not trained.

This can lead to strange behavior like

  • atoms crashing into each other

  • extremely high temperatures

The active learning workflow will try to detect these events and discard any subsequent structures.

If you train a ParAMS ML Committee (MachineLearning%CommitteeSize > 1), the ML model will also return an estimated uncertainty of predicted energies and forces.

You can also set a threshold for these uncertainties, such that if they are exceeded the MD simulation immediately stops, even before the end of the active learning step. You can thus choose to use the predicted uncertainties to decide when to stop the simulation, and use structures with high uncertainty for the training set. This method can be used in addition to active learning step division.

Criterion

Implementation

Temperature

inside active learning workflow

Distance

AMS Exit Condition

Energy uncertainty

AMS Exit Condition

Forces uncertainty

AMS Exit Condition

Note

If a “reasonable simulation criterion” is exceeded, this will never count as a successful step/attempt.

It will always lead to a retraining of the model and an increase of the attempt number, even if MaxAttemptsPerStep is exceeded.

ActiveLearning
   ReasonableSimulationCriteria
      Distance
         Enabled Yes/No
         MinValue float
      End
      EnergyUncertainty
         Enabled Yes/No
         MaxValue float
         Normalization float
      End
      GradientsUncertainty
         Enabled Yes/No
         MaxValue float
      End
      Temperature
         Enabled Yes/No
         MaxValue float
      End
   End
End
ActiveLearning
Type:

Block

Description:

Settings for Active Learning

ReasonableSimulationCriteria
Type:

Block

Description:

Criteria for determining whether a simulation is reasonable. If any of the criteria are exceeded, this will be reported as ‘ENERGY_UNCERTAINTY’, ‘TEMPERATURE’, etc., with capital letters in the output. If a simulation is unreasonable, it will never lead to an increase of the Step, even if the number of attempts exceeds MaxAttemptsPerStep.

Distance
Type:

Block

Description:

Stop the simulation if any interatomic distance is smaller than the specified value.

Enabled
Type:

Bool

Default value:

Yes

Description:

Stop the simulation if any interatomic distance is smaller than the specified value.

MinValue
Type:

Float

Default value:

0.6

Unit:

angstrom

GUI name:

Minimum

Description:

Minimum allowed interatomic distance.

EnergyUncertainty
Type:

Block

Description:

Stop the simulation if the uncertainty in the energy is too high. Currently only applicable when training committees.

Enabled
Type:

Bool

Default value:

No

Description:

Stop the simulation if the uncertainty in the energy is too high. Currently only applicable when training committees. If CommitteeSize = 1 then this keyword has no effect.

MaxValue
Type:

Float

Default value:

0.015

Unit:

eV

GUI name:

Maximum

Description:

Threshold for allowed [energy uncertainty divided by Normalization].

Normalization
Type:

Float

Description:

Normalize (divide) the energy uncertainty by this number before comparing to the specified threshold. If not specified, it will become the number of atoms.

GradientsUncertainty
Type:

Block

Description:

Stop the simulation if the uncertainty in the gradients (forces) is too high. Currently only applicable when training committees.

Enabled
Type:

Bool

Default value:

No

Description:

Stop the simulation if the uncertainty in the gradients (forces) is too high. Currently only applicable when training committees. If CommitteeSize = 1 then this keyword has no effect.

MaxValue
Type:

Float

Default value:

0.5

Unit:

eV/angstrom

GUI name:

Maximum

Description:

Maximum allowed gradients (forces) uncertainty.

Temperature
Type:

Block

Description:

Discard all frames after the temperature has reached the specified value.

Enabled
Type:

Bool

Default value:

Yes

Description:

Discard all frames after the temperature has reached the specified value.

MaxValue
Type:

Float

Default value:

5000.0

Unit:

K

GUI name:

Maximum

Description:

Maximum allowed temperature

Output to save

The active learning workflow produces many directories containing reference calculations, MD simulations, and ParAMS training. You can choose how much output to save.

By default, the workflow only keeps the directories it needs to keep going. This always includes

  • the entire training and validation sets, and

  • the MD trajectory from the beginning of the workflow.

By default, the reference calculation directories are not saved unless the reference calculation fails.

ActiveLearning
   Save
      ReferenceCalculations [None | All]
      ReferenceData [Latest | All]
      TrainingDirectories [Latest | All]
      Trajectories [Latest | All]
   End
End
ActiveLearning
Type:

Block

Description:

Settings for Active Learning

Save
Type:

Block

Description:

The files/directories on disk to keep. If you set these options to All, a lot of output will be created. This output is usually not necessary but can be used for debugging purposes, or to better understand what the workflow is doing.

ReferenceCalculations
Type:

Multiple Choice

Default value:

None

Options:

[None, All]

Description:

The reference calculation directories (initial_reference_calculations or stepX_attemptY_ref_calcZ) including the original input and output. These directories may take up a lot of disk space and are not kept by default. Enable this option if you need to investigate why reference calculations fail (incorrect input, SCF convergence problems, …), or if you want to keep them for some other reason. Note: The output used for parametrization (energy, forces) is always stored in the ReferenceData (training and validation sets).

ReferenceData
Type:

Multiple Choice

Default value:

Latest

Options:

[Latest, All]

Description:

The reference data directories (stepX_attemptY_reference_data) containing the training and validation sets in ParAMS .yaml format (and ASE .xyz format). These can be opened in the ParAMS GUI or used as input for ParAMS.

TrainingDirectories
Type:

Multiple Choice

Default value:

Latest

Options:

[Latest, All]

Description:

The ParAMS training directories (stepX_attemptY_training).

Trajectories
Type:

Multiple Choice

Default value:

Latest

Options:

[Latest, All]

Description:

The MD trajectory calculation directories (stepX_attemptY_simulation) using the trained ML potential. Note: the trajectories in these directories are the entire trajectories from the beginning of the simulation.

At workflow end: retrain model, rerun simulation

Retrain model

After the final active learning step, you have the option to retrain the model using all reference data.

This may be useful to not “waste” reference calculations that have been performed but not used for training.

Example: if the the last 3 active learning steps are successful at the first attempt, then the workflow will have run 3 reference calculations (for the evaluation of the success criteria) that have not been used for training or validation.

The downside of retraining the model is that you may end up with a model that would have failed the success criteria!

By default, the model is not automatically retrained.

Rerun simulation (final production simulation)

After the final active learning step is successful, you can rerun the entire MD simulation from scratch using the final model parameters.

This will give you an MD trajectory with consistent sampling frequency and calculated using a single potential energy surface.

It is run in a directory called final_production_simulation, and replaces the ams.rkf file in the main results directory.

The reasonable simulation criteria are not applied to the final production simulation.

AtEnd input

ActiveLearning
   AtEnd
      RerunSimulation Yes/No
      RetrainModel Yes/No
   End
End
ActiveLearning
Type:

Block

Description:

Settings for Active Learning

AtEnd
Type:

Block

Description:

What to do at the end of the active learning loop.

RerunSimulation
Type:

Bool

Default value:

Yes

Description:

Rerun the MD simulation (folder: final_production_simulation) using the last set of parameters. This guarantees that the entire trajectory is calculated using the same model / potential energy surface, and that the trajectory has a consistent sampling frequency. This means that it can be used with all MD postanalysis tools.

RetrainModel
Type:

Bool

Default value:

No

Description:

Train a final model (folder: final_training) using all reference (training and validation) data, including any reference calculations that have not yet been trained to.