Best practices for machine-learned force fields - Revision history

Csheldon: /* Ab-initio calculation setup */

2026-03-24T13:38:57Z

Ab-initio calculation setup

← Older revision		Revision as of 13:38, 24 March 2026
Line 46:		Line 46:
	* It is important to learn the exact forces. To do this, the electronic minimization has to be checked for convergence. These checks may include, for example, the number of k-points in the {{FILE\|KPOINTS}} file, the plane wave limit ({{TAG\|ENCUT}}), the electronic minimization algorithm, etc.		* It is important to learn the exact forces. To do this, the electronic minimization has to be checked for convergence. These checks may include, for example, the number of k-points in the {{FILE\|KPOINTS}} file, the plane wave limit ({{TAG\|ENCUT}}), the electronic minimization algorithm, etc.
	* Turn off symmetry as for standard molecular dynamics runs ({{TAG\|ISYM}}=0).		* Turn off symmetry as for standard molecular dynamics runs ({{TAG\|ISYM}}=0).
	* For simulations without a fixed grid (NpT), the cutoff for plane waves {{TAG\|ENCUT}} must be set at least 30 percent higher than for fixed volume calculations. Also, it is good to restart frequently ({{TAG\|ML_MODE}}=TRAIN with existing {{FILE\|ML_AB}} file in working directory) to reinitialize the [[Projector-augmented-wave formalism\|PAW]] basis of KS orbitals and avoid [[~~Energy vs volume Volume relaxations and Pulay stress#Pulay stress\|~~Pulay stress]].		* For simulations without a fixed grid (NpT), the cutoff for plane waves {{TAG\|ENCUT}} must be set at least 30 percent higher than for fixed volume calculations. Also, it is good to restart frequently ({{TAG\|ML_MODE}}=TRAIN with existing {{FILE\|ML_AB}} file in working directory) to reinitialize the [[Projector-augmented-wave formalism\|PAW]] basis of KS orbitals and avoid [[Pulay stress]].
	{{NB\|warning\|It is very important to '''not''' change the ab-initio settings in the {{FILE\|INCAR}} file between training from scratch and continuing training. Likewise, the {{FILE\|POTCAR}} file is '''not''' allowed to be changed when resuming training.}}		{{NB\|warning\|It is very important to '''not''' change the ab-initio settings in the {{FILE\|INCAR}} file between training from scratch and continuing training. Likewise, the {{FILE\|POTCAR}} file is '''not''' allowed to be changed when resuming training.}}

	==== Molecular dynamics setup ====		==== Molecular dynamics setup ====

Karsai: /* Performance */

2026-03-20T07:56:33Z

Performance

← Older revision		Revision as of 07:56, 20 March 2026
Line 215:		Line 215:
	The machine learning code is parallelized using MPI.		The machine learning code is parallelized using MPI.
	It is most efficient if scaLAPACK is used since storing (and working on) large matrices, in particular the design matrix, will then be distributed over the MPI ranks. However, a LAPACK-only version exists as well. In the latter case, only a few matrices are stored in a distributed fashion, so due to the high memory demand, the LAPACK version is not feasible for "realistic" systems.		It is most efficient if scaLAPACK is used since storing (and working on) large matrices, in particular the design matrix, will then be distributed over the MPI ranks. However, a LAPACK-only version exists as well. In the latter case, only a few matrices are stored in a distributed fashion, so due to the high memory demand, the LAPACK version is not feasible for "realistic" systems.

	{{NB\|warning\|When compiling with shared memory MPI support (-Duse_shmem), it is utterly important to pin the MPI ranks to the physical cores of the node. For guidance on how to understand the hardware topology of your system and how to correctly set up rank pinning accordingly, refer to [[Optimizing the parallelization#Understanding the hardware\|here]].}}		{{NB\|warning\|When compiling with shared memory MPI support (-Duse_shmem), it is utterly important to pin the MPI ranks to the physical cores of the node. For guidance on how to understand the hardware topology of your system and how to correctly set up rank pinning accordingly, refer to [[Optimizing the parallelization#Understanding the hardware\|here]].}}

Karsai: /* Performance */

2026-03-20T07:56:26Z

Performance

← Older revision		Revision as of 07:56, 20 March 2026
Line 215:		Line 215:
	The machine learning code is parallelized using MPI.		The machine learning code is parallelized using MPI.
	It is most efficient if scaLAPACK is used since storing (and working on) large matrices, in particular the design matrix, will then be distributed over the MPI ranks. However, a LAPACK-only version exists as well. In the latter case, only a few matrices are stored in a distributed fashion, so due to the high memory demand, the LAPACK version is not feasible for "realistic" systems.		It is most efficient if scaLAPACK is used since storing (and working on) large matrices, in particular the design matrix, will then be distributed over the MPI ranks. However, a LAPACK-only version exists as well. In the latter case, only a few matrices are stored in a distributed fashion, so due to the high memory demand, the LAPACK version is not feasible for "realistic" systems.

			{{NB\|warning\|When compiling with shared memory MPI support (-Duse_shmem), it is utterly important to pin the MPI ranks to the physical cores of the node. For guidance on how to understand the hardware topology of your system and how to correctly set up rank pinning accordingly, refer to [[Optimizing the parallelization#Understanding the hardware\|here]].}}

	=== Computational efficiency in production runs ===		=== Computational efficiency in production runs ===

Karsai: /* Computational efficiency in production runs */

2026-03-20T07:56:05Z

Computational efficiency in production runs

← Older revision		Revision as of 07:56, 20 March 2026
Line 217:		Line 217:

	=== Computational efficiency in production runs ===		=== Computational efficiency in production runs ===
	{{NB\|warning\|When compiling with shared memory MPI support (-Duse_shmem), it is utterly important to pin the MPI ranks to the physical cores of the node. For guidance on how to understand the hardware topology of your system and how to correctly set up rank pinning accordingly, refer to [[Optimizing the parallelization#Understanding the hardware\|here]].}}
	{{NB\|tip\|The most important step towards efficient production runs is to apply {{TAG\|ML_MODE\|refit\|color=blue}} in order to obtain an {{FILE\|ML_FFN}} force field file which supports the fast prediction mode (available as of {{VASP}} 6.4.0). Please have a look at step 4 in the [[Machine_learning_force_field_calculations:_Basics#Step-by-step_instructions\|basic step-by-step instructions]]. Whether your force field file supports fast prediction can be checked in the [[ML_FFN\|file header]] (independent of running {{VASP}}) or in the {{FILE\|ML_LOGFILE}} (after starting {{TAG\|ML_MODE\|run\|color=blue}}). Speedups with respect to a force field file without support for fast prediction mode are typically of the order 20 to 100.}}		{{NB\|tip\|The most important step towards efficient production runs is to apply {{TAG\|ML_MODE\|refit\|color=blue}} in order to obtain an {{FILE\|ML_FFN}} force field file which supports the fast prediction mode (available as of {{VASP}} 6.4.0). Please have a look at step 4 in the [[Machine_learning_force_field_calculations:_Basics#Step-by-step_instructions\|basic step-by-step instructions]]. Whether your force field file supports fast prediction can be checked in the [[ML_FFN\|file header]] (independent of running {{VASP}}) or in the {{FILE\|ML_LOGFILE}} (after starting {{TAG\|ML_MODE\|run\|color=blue}}). Speedups with respect to a force field file without support for fast prediction mode are typically of the order 20 to 100.}}
	This section addresses challenges encountered in production runs utilizing the force field ({{TAG\|ML_MODE}}=''RUN'') with the fast version (requiring prior refitting using {{TAG\|ML_MODE}}=''refit''). The time required to evaluate a single step of this force field typically matches the duration needed to write results to files. Furthermore, the file-writing process operates solely in a serial manner. As the number of computational cores increases, the overall time for force field evaluation decreases, while the file-writing time remains constant. Therefore, optimizing performance necessitates adjusting the output frequency. The following flags can be used for that:		This section addresses challenges encountered in production runs utilizing the force field ({{TAG\|ML_MODE}}=''RUN'') with the fast version (requiring prior refitting using {{TAG\|ML_MODE}}=''refit''). The time required to evaluate a single step of this force field typically matches the duration needed to write results to files. Furthermore, the file-writing process operates solely in a serial manner. As the number of computational cores increases, the overall time for force field evaluation decreases, while the file-writing time remains constant. Therefore, optimizing performance necessitates adjusting the output frequency. The following flags can be used for that:

Karsai: /* Performance */

2026-03-20T07:55:40Z

Performance

← Older revision		Revision as of 07:55, 20 March 2026
Line 217:		Line 217:

	=== Computational efficiency in production runs ===		=== Computational efficiency in production runs ===
			{{NB\|warning\|When compiling with shared memory MPI support (-Duse_shmem), it is utterly important to pin the MPI ranks to the physical cores of the node. For guidance on how to understand the hardware topology of your system and how to correctly set up rank pinning accordingly, refer to [[Optimizing the parallelization#Understanding the hardware\|here]].}}
	{{NB\|tip\|The most important step towards efficient production runs is to apply {{TAG\|ML_MODE\|refit\|color=blue}} in order to obtain an {{FILE\|ML_FFN}} force field file which supports the fast prediction mode (available as of {{VASP}} 6.4.0). Please have a look at step 4 in the [[Machine_learning_force_field_calculations:_Basics#Step-by-step_instructions\|basic step-by-step instructions]]. Whether your force field file supports fast prediction can be checked in the [[ML_FFN\|file header]] (independent of running {{VASP}}) or in the {{FILE\|ML_LOGFILE}} (after starting {{TAG\|ML_MODE\|run\|color=blue}}). Speedups with respect to a force field file without support for fast prediction mode are typically of the order 20 to 100.}}		{{NB\|tip\|The most important step towards efficient production runs is to apply {{TAG\|ML_MODE\|refit\|color=blue}} in order to obtain an {{FILE\|ML_FFN}} force field file which supports the fast prediction mode (available as of {{VASP}} 6.4.0). Please have a look at step 4 in the [[Machine_learning_force_field_calculations:_Basics#Step-by-step_instructions\|basic step-by-step instructions]]. Whether your force field file supports fast prediction can be checked in the [[ML_FFN\|file header]] (independent of running {{VASP}}) or in the {{FILE\|ML_LOGFILE}} (after starting {{TAG\|ML_MODE\|run\|color=blue}}). Speedups with respect to a force field file without support for fast prediction mode are typically of the order 20 to 100.}}
	This section addresses challenges encountered in production runs utilizing the force field ({{TAG\|ML_MODE}}=''RUN'') with the fast version (requiring prior refitting using {{TAG\|ML_MODE}}=''refit''). The time required to evaluate a single step of this force field typically matches the duration needed to write results to files. Furthermore, the file-writing process operates solely in a serial manner. As the number of computational cores increases, the overall time for force field evaluation decreases, while the file-writing time remains constant. Therefore, optimizing performance necessitates adjusting the output frequency. The following flags can be used for that:		This section addresses challenges encountered in production runs utilizing the force field ({{TAG\|ML_MODE}}=''RUN'') with the fast version (requiring prior refitting using {{TAG\|ML_MODE}}=''refit''). The time required to evaluate a single step of this force field typically matches the duration needed to write results to files. Furthermore, the file-writing process operates solely in a serial manner. As the number of computational cores increases, the overall time for force field evaluation decreases, while the file-writing time remains constant. Therefore, optimizing performance necessitates adjusting the output frequency. The following flags can be used for that:

Csheldon: /* Spilling factor: error estimates during production runs */

2026-02-10T13:14:53Z

Spilling factor: error estimates during production runs

← Older revision		Revision as of 13:14, 10 February 2026
Line 183:		Line 183:
	Using the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] one can measure the error during the production runs. To do so one has to set the {{TAG\|ML_ESTBLOCK\|0\|op=>}} in the {{TAG\|INCAR}} file (the default value is 0). This tag controls after how many molecular dynamics steps the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] is calculated. The calculation of the spilling factor scales quadratically with the number of local reference configurations and linearly with the number of species. So for force fields containing many species and/or local reference configurations, the evaluation time of the spilling factor becomes of the order of magnitude or more as the evaluation of a single force field step. Since it is enough to monitor the error only after the ions moved several MD steps, the total time consumed by evaluating the spilling factor can become insignificantly compared to the total time. So in long molecular dynamics calculations, we recommend using at least {{TAG\|ML_ESTBLOCK\|20-100}}.		Using the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] one can measure the error during the production runs. To do so one has to set the {{TAG\|ML_ESTBLOCK\|0\|op=>}} in the {{TAG\|INCAR}} file (the default value is 0). This tag controls after how many molecular dynamics steps the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] is calculated. The calculation of the spilling factor scales quadratically with the number of local reference configurations and linearly with the number of species. So for force fields containing many species and/or local reference configurations, the evaluation time of the spilling factor becomes of the order of magnitude or more as the evaluation of a single force field step. Since it is enough to monitor the error only after the ions moved several MD steps, the total time consumed by evaluating the spilling factor can become insignificantly compared to the total time. So in long molecular dynamics calculations, we recommend using at least {{TAG\|ML_ESTBLOCK\|20-100}}.

	The [[Machine learning force field: Theory#Spilling factor\|spilling factor]] measures the similarity of the local environment of each atom in the current structure to that of the local reference configurations of the force field. The values of the spilling factor are in the range <math>[0,1]</math>. lf the atomic environment is "properly" represented by the local reference configurations the spilling factor approaches 0. Vice versa the spilling factor approaches quickly 1, meaning that the force field is probably extrapolating. Molecular dynamics trajectories where the spilling factor is most of the time 1 can still lead to good results, but the calculations should be cautiously used.		The [[Machine learning force field: Theory#Spilling factor\|spilling factor]] measures the similarity of the local environment of each atom in the current structure to that of the local reference configurations of the force field. The values of the spilling factor are in the range <math>[0,1]</math>. lf the atomic environment is "properly" represented by the local reference configurations the spilling factor approaches 0. Vice versa the spilling factor approaches quickly 1, meaning that the force field is probably extrapolating. Molecular dynamics trajectories where the spilling factor is most of the time 1 can still lead to good results, but the calculations should be cautiously used. <!--You can systematically improve the force field by extracting the structures where the spilling factor is larger than 1 and using [[INTERACTIVE \| interactive mode]] to continue [[Construction:Using metadynamics to train a machine-learned force field \| training the force field specifically on these structures]].-->

	Besides being able to monitor the accuracy during the production runs one can also use the spilling factor to assess the accuracy of the force field for given test sets (i.e. structures chosen from ensembles at different temperatures) like in the traditional way, where test structures are picked out and ab initio calculations have to be carried out for each structure. Using the spilling factor the error is directly assessed without the need for ab initio calculations making the procedure orders of magnitude faster and easier to handle (no evaluation script needed). Nevertheless, if one wants to measure the true error on a test set we have described how to [[Best practices for machine-learned force fields#Test errors\|below]].		Besides being able to monitor the accuracy during the production runs one can also use the spilling factor to assess the accuracy of the force field for given test sets (i.e. structures chosen from ensembles at different temperatures) like in the traditional way, where test structures are picked out and ab initio calculations have to be carried out for each structure. Using the spilling factor the error is directly assessed without the need for ab initio calculations making the procedure orders of magnitude faster and easier to handle (no evaluation script needed). Nevertheless, if one wants to measure the true error on a test set we have described how to [[Best practices for machine-learned force fields#Test errors\|below]].

Singraber: /* Testing and application */

2026-02-05T08:11:30Z

Testing and application

← Older revision		Revision as of 08:11, 5 February 2026
Line 181:		Line 181:

	=== Spilling factor: error estimates during production runs===		=== Spilling factor: error estimates during production runs===
	Using the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] one can measure the error during the production runs. To do so one has to set the {{TAG\|~~ML_IERR~~\|0\|op=>}} in the {{TAG\|INCAR}} file (the default value is 0). This tag controls after how many molecular dynamics steps the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] is calculated. The calculation of the spilling factor scales quadratically with the number of local reference configurations and linearly with the number of species. So for force fields containing many species and/or local reference configurations, the evaluation time of the spilling factor becomes of the order of magnitude or more as the evaluation of a single force field step. Since it is enough to monitor the error only after the ions moved several MD steps, the total time consumed by evaluating the spilling factor can become insignificantly compared to the total time. So in long molecular dynamics calculations, we recommend using at least {{TAG\|~~ML_IERR}}=~~20-100.		Using the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] one can measure the error during the production runs. To do so one has to set the {{TAG\|ML_ESTBLOCK\|0\|op=>}} in the {{TAG\|INCAR}} file (the default value is 0). This tag controls after how many molecular dynamics steps the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] is calculated. The calculation of the spilling factor scales quadratically with the number of local reference configurations and linearly with the number of species. So for force fields containing many species and/or local reference configurations, the evaluation time of the spilling factor becomes of the order of magnitude or more as the evaluation of a single force field step. Since it is enough to monitor the error only after the ions moved several MD steps, the total time consumed by evaluating the spilling factor can become insignificantly compared to the total time. So in long molecular dynamics calculations, we recommend using at least {{TAG\|ML_ESTBLOCK\|20-100}}.

	The [[Machine learning force field: Theory#Spilling factor\|spilling factor]] measures the similarity of the local environment of each atom in the current structure to that of the local reference configurations of the force field. The values of the spilling factor are in the range <math>[0,1]</math>. lf the atomic environment is "properly" represented by the local reference configurations the spilling factor approaches 0. Vice versa the spilling factor approaches quickly 1, meaning that the force field is probably extrapolating. Molecular dynamics trajectories where the spilling factor is most of the time 1 can still lead to good results, but the calculations should be cautiously used.		The [[Machine learning force field: Theory#Spilling factor\|spilling factor]] measures the similarity of the local environment of each atom in the current structure to that of the local reference configurations of the force field. The values of the spilling factor are in the range <math>[0,1]</math>. lf the atomic environment is "properly" represented by the local reference configurations the spilling factor approaches 0. Vice versa the spilling factor approaches quickly 1, meaning that the force field is probably extrapolating. Molecular dynamics trajectories where the spilling factor is most of the time 1 can still lead to good results, but the calculations should be cautiously used.

Csheldon: /* Retraining with hyper-parameter optimization */

2025-11-14T11:21:36Z

Retraining with hyper-parameter optimization

← Older revision		Revision as of 11:21, 14 November 2025
Line 73:		Line 73:
	After you have collected a sufficient number of local atomic reference configurations, as described in Training from scratch and Continuation training, it is recommended to		After you have collected a sufficient number of local atomic reference configurations, as described in Training from scratch and Continuation training, it is recommended to
	optimize the parameters for your force field. This will result in		optimize the parameters for your force field. This will result in
	lower training and test set errors. The reference configurations selected in the {{FILE\|ML_AB}} will not be updated. To perform a hyperparameter search, {{TAG\|ML_MODE}}=REFIT must be set in the {{FILE\|INCAR}} file and a {{FILE\|ML_AB}} must exist in the working directory. By setting {{TAG\|ML_MODE}}=REFIT, {{VASP}} automatically selects {{TAG\|ML_IALGO_LINREG}}=4, which performs a regularized SVD to find the appropriate weights <math> \mathbf{w} </math> (see [[Machine_learning_force_field:_Theory#Matrix_vector_form_of_linear_equations\|here]] for the definition). It is favorable to enter refit mode and tune the hyperparameters to improve the fitting error, which can be found in the {{FILE\|ML_LOGFILE}} under the description '''ERR'''. To tune the hyperparameters, set the desired parameter in the {{FILE\|INCAR}} file, then run {{VASP}} and check the error in the {{FILE\|ML_LOGFILE}}. For more information on extracting errors from the {{FILE\|ML_LOGFILE}}, see [[~~Best Practices for Machine~~-~~Learning Force Fields~~#~~Monitoring~~\|here]]. Adjusting the following parameters may improve the quality of the force-fields:		lower training and test set errors. The reference configurations selected in the {{FILE\|ML_AB}} will not be updated. To perform a hyperparameter search, {{TAG\|ML_MODE}}=REFIT must be set in the {{FILE\|INCAR}} file and a {{FILE\|ML_AB}} must exist in the working directory. By setting {{TAG\|ML_MODE}}=REFIT, {{VASP}} automatically selects {{TAG\|ML_IALGO_LINREG}}=4, which performs a regularized SVD to find the appropriate weights <math> \mathbf{w} </math> (see [[Machine_learning_force_field:_Theory#Matrix_vector_form_of_linear_equations\|here]] for the definition). It is favorable to enter refit mode and tune the hyperparameters to improve the fitting error, which can be found in the {{FILE\|ML_LOGFILE}} under the description '''ERR'''. To tune the hyperparameters, set the desired parameter in the {{FILE\|INCAR}} file, then run {{VASP}} and check the error in the {{FILE\|ML_LOGFILE}}. For more information on extracting errors from the {{FILE\|ML_LOGFILE}}, see [[Best_practices_for_machine-learned_force_fields#Monitoring_on-the-fly_learning\|here]]. Adjusting the following parameters may improve the quality of the force-fields:
	* Adjusting the cutoff radius for the angular and radial descriptor by adjusting {{TAG\|ML_RCUT2}} and {{TAG\|ML_RCUT1}}.		* Adjusting the cutoff radius for the angular and radial descriptor by adjusting {{TAG\|ML_RCUT2}} and {{TAG\|ML_RCUT1}}.
	* Matching the number of radial and angular basis functions with {{TAG\|ML_MRB1}} and {{TAG\|ML_MRB2}}.		* Matching the number of radial and angular basis functions with {{TAG\|ML_MRB1}} and {{TAG\|ML_MRB2}}.

Csheldon: /* Retraining with hyper-parameter optimization */

2025-11-14T11:20:03Z

Retraining with hyper-parameter optimization

← Older revision		Revision as of 11:20, 14 November 2025
Line 73:		Line 73:
	After you have collected a sufficient number of local atomic reference configurations, as described in Training from scratch and Continuation training, it is recommended to		After you have collected a sufficient number of local atomic reference configurations, as described in Training from scratch and Continuation training, it is recommended to
	optimize the parameters for your force field. This will result in		optimize the parameters for your force field. This will result in
	lower training and test set errors. The reference configurations selected in the {{FILE\|ML_AB}} will not be updated. To perform a hyperparameter search, {{TAG\|ML_MODE}}=REFIT must be set in the {{FILE\|INCAR}} file and a {{FILE\|ML_AB}} must exist in the working directory. By setting {{TAG\|ML_MODE}}=REFIT, {{VASP}} automatically selects {{TAG\|ML_IALGO_LINREG}}=4, which performs a regularized SVD to find the appropriate weights <math> \mathbf{w} </math> (see [[~~Machine Learning Force Field~~: ~~Theory~~#~~Matrix Vector Shape of Linear Equations~~\|here]] for the definition). It is favorable to enter refit mode and tune the hyperparameters to improve the fitting error, which can be found in the {{FILE\|ML_LOGFILE}} under the description '''ERR'''. To tune the hyperparameters, set the desired parameter in the {{FILE\|INCAR}} file, then run {{VASP}} and check the error in the {{FILE\|ML_LOGFILE}}. For more information on extracting errors from the {{FILE\|ML_LOGFILE}}, see [[Best Practices for Machine-Learning Force Fields#Monitoring\|here]]. Adjusting the following parameters may improve the quality of the force-fields:		lower training and test set errors. The reference configurations selected in the {{FILE\|ML_AB}} will not be updated. To perform a hyperparameter search, {{TAG\|ML_MODE}}=REFIT must be set in the {{FILE\|INCAR}} file and a {{FILE\|ML_AB}} must exist in the working directory. By setting {{TAG\|ML_MODE}}=REFIT, {{VASP}} automatically selects {{TAG\|ML_IALGO_LINREG}}=4, which performs a regularized SVD to find the appropriate weights <math> \mathbf{w} </math> (see [[Machine_learning_force_field:_Theory#Matrix_vector_form_of_linear_equations\|here]] for the definition). It is favorable to enter refit mode and tune the hyperparameters to improve the fitting error, which can be found in the {{FILE\|ML_LOGFILE}} under the description '''ERR'''. To tune the hyperparameters, set the desired parameter in the {{FILE\|INCAR}} file, then run {{VASP}} and check the error in the {{FILE\|ML_LOGFILE}}. For more information on extracting errors from the {{FILE\|ML_LOGFILE}}, see [[Best Practices for Machine-Learning Force Fields#Monitoring\|here]]. Adjusting the following parameters may improve the quality of the force-fields:
	* Adjusting the cutoff radius for the angular and radial descriptor by adjusting {{TAG\|ML_RCUT2}} and {{TAG\|ML_RCUT1}}.		* Adjusting the cutoff radius for the angular and radial descriptor by adjusting {{TAG\|ML_RCUT2}} and {{TAG\|ML_RCUT1}}.
	* Matching the number of radial and angular basis functions with {{TAG\|ML_MRB1}} and {{TAG\|ML_MRB2}}.		* Matching the number of radial and angular basis functions with {{TAG\|ML_MRB1}} and {{TAG\|ML_MRB2}}.

Huebsch at 11:35, 24 October 2025

2025-10-24T11:35:18Z

← Older revision		Revision as of 11:35, 24 October 2025
Line 37:		Line 37:
	The main disadvantage is decreased computational efficiency. The computational cost scales quadratically with the number of species. Using the [[Best practices for machine-learned force fields#Descriptor reduction in production runs\|reduced descriptor]] ({{TAGDEF\|ML_DESC_TYPE\|1}}) can reduce this to linear scaling for major parts of the code, but even perfect linear scaling introduces noticeable overhead when increasing the number of species.		The main disadvantage is decreased computational efficiency. The computational cost scales quadratically with the number of species. Using the [[Best practices for machine-learned force fields#Descriptor reduction in production runs\|reduced descriptor]] ({{TAGDEF\|ML_DESC_TYPE\|1}}) can reduce this to linear scaling for major parts of the code, but even perfect linear scaling introduces noticeable overhead when increasing the number of species.
	{{NB\|warning\|It is not possible to give the same name to different groups of atoms in the {{FILE\|POSCAR}} file and the names are restricted to two characters.}}		{{NB\|warning\|It is not possible to give the same name to different groups of atoms in the {{FILE\|POSCAR}} file and the names are restricted to two characters.}}
	~~'''~~ Ab-initio calculation setup ~~'''~~		==== Ab-initio calculation setup ====

	The training mode requires {{VASP}} to perform ab-initio calculations, so the first step is to set up the [[:Category:Electronic minimization\|electronic minimization]] scheme.		The training mode requires {{VASP}} to perform ab-initio calculations, so the first step is to set up the [[:Category:Electronic minimization\|electronic minimization]] scheme.
Line 48:		Line 48:
	* For simulations without a fixed grid (NpT), the cutoff for plane waves {{TAG\|ENCUT}} must be set at least 30 percent higher than for fixed volume calculations. Also, it is good to restart frequently ({{TAG\|ML_MODE}}=TRAIN with existing {{FILE\|ML_AB}} file in working directory) to reinitialize the [[Projector-augmented-wave formalism\|PAW]] basis of KS orbitals and avoid [[Energy vs volume Volume relaxations and Pulay stress#Pulay stress\|Pulay stress]].		* For simulations without a fixed grid (NpT), the cutoff for plane waves {{TAG\|ENCUT}} must be set at least 30 percent higher than for fixed volume calculations. Also, it is good to restart frequently ({{TAG\|ML_MODE}}=TRAIN with existing {{FILE\|ML_AB}} file in working directory) to reinitialize the [[Projector-augmented-wave formalism\|PAW]] basis of KS orbitals and avoid [[Energy vs volume Volume relaxations and Pulay stress#Pulay stress\|Pulay stress]].
	{{NB\|warning\|It is very important to '''not''' change the ab-initio settings in the {{FILE\|INCAR}} file between training from scratch and continuing training. Likewise, the {{FILE\|POTCAR}} file is '''not''' allowed to be changed when resuming training.}}		{{NB\|warning\|It is very important to '''not''' change the ab-initio settings in the {{FILE\|INCAR}} file between training from scratch and continuing training. Likewise, the {{FILE\|POTCAR}} file is '''not''' allowed to be changed when resuming training.}}
	~~'''~~ Molecular dynamics setup ~~'''~~		==== Molecular dynamics setup ====

	After the forces are obtained from electronic minimization by the [[Hellmann-Feynman forces\|Hellmann-Feynman Theorem]], {{VASP}} must propagate the ions to obtain a new configuration in phase space. For the molecular dynamics part, familiarity with setting up {{TAG\|molecular dynamics}} runs is beneficial. In addition, we recommend the following settings in the molecular dynamics part:		After the forces are obtained from electronic minimization by the [[Hellmann-Feynman forces\|Hellmann-Feynman Theorem]], {{VASP}} must propagate the ions to obtain a new configuration in phase space. For the molecular dynamics part, familiarity with setting up {{TAG\|molecular dynamics}} runs is beneficial. In addition, we recommend the following settings in the molecular dynamics part:

← Older revision		Revision as of 13:38, 24 March 2026
Line 46:		Line 46:
	* It is important to learn the exact forces. To do this, the electronic minimization has to be checked for convergence. These checks may include, for example, the number of k-points in the {{FILE\|KPOINTS}} file, the plane wave limit ({{TAG\|ENCUT}}), the electronic minimization algorithm, etc.		* It is important to learn the exact forces. To do this, the electronic minimization has to be checked for convergence. These checks may include, for example, the number of k-points in the {{FILE\|KPOINTS}} file, the plane wave limit ({{TAG\|ENCUT}}), the electronic minimization algorithm, etc.
	* Turn off symmetry as for standard molecular dynamics runs ({{TAG\|ISYM}}=0).		* Turn off symmetry as for standard molecular dynamics runs ({{TAG\|ISYM}}=0).
	* For simulations without a fixed grid (NpT), the cutoff for plane waves {{TAG\|ENCUT}} must be set at least 30 percent higher than for fixed volume calculations. Also, it is good to restart frequently ({{TAG\|ML_MODE}}=TRAIN with existing {{FILE\|ML_AB}} file in working directory) to reinitialize the [[Projector-augmented-wave formalism\|PAW]] basis of KS orbitals and avoid [[~~Energy vs volume Volume relaxations and Pulay stress#Pulay stress\|~~Pulay stress]].		* For simulations without a fixed grid (NpT), the cutoff for plane waves {{TAG\|ENCUT}} must be set at least 30 percent higher than for fixed volume calculations. Also, it is good to restart frequently ({{TAG\|ML_MODE}}=TRAIN with existing {{FILE\|ML_AB}} file in working directory) to reinitialize the [[Projector-augmented-wave formalism\|PAW]] basis of KS orbitals and avoid [[Pulay stress]].
	{{NB\|warning\|It is very important to '''not''' change the ab-initio settings in the {{FILE\|INCAR}} file between training from scratch and continuing training. Likewise, the {{FILE\|POTCAR}} file is '''not''' allowed to be changed when resuming training.}}		{{NB\|warning\|It is very important to '''not''' change the ab-initio settings in the {{FILE\|INCAR}} file between training from scratch and continuing training. Likewise, the {{FILE\|POTCAR}} file is '''not''' allowed to be changed when resuming training.}}

	==== Molecular dynamics setup ====		==== Molecular dynamics setup ====

← Older revision		Revision as of 13:14, 10 February 2026
Line 183:		Line 183:
	Using the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] one can measure the error during the production runs. To do so one has to set the {{TAG\|ML_ESTBLOCK\|0\|op=>}} in the {{TAG\|INCAR}} file (the default value is 0). This tag controls after how many molecular dynamics steps the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] is calculated. The calculation of the spilling factor scales quadratically with the number of local reference configurations and linearly with the number of species. So for force fields containing many species and/or local reference configurations, the evaluation time of the spilling factor becomes of the order of magnitude or more as the evaluation of a single force field step. Since it is enough to monitor the error only after the ions moved several MD steps, the total time consumed by evaluating the spilling factor can become insignificantly compared to the total time. So in long molecular dynamics calculations, we recommend using at least {{TAG\|ML_ESTBLOCK\|20-100}}.		Using the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] one can measure the error during the production runs. To do so one has to set the {{TAG\|ML_ESTBLOCK\|0\|op=>}} in the {{TAG\|INCAR}} file (the default value is 0). This tag controls after how many molecular dynamics steps the [[Machine learning force field: Theory#Spilling factor\|spilling factor]] is calculated. The calculation of the spilling factor scales quadratically with the number of local reference configurations and linearly with the number of species. So for force fields containing many species and/or local reference configurations, the evaluation time of the spilling factor becomes of the order of magnitude or more as the evaluation of a single force field step. Since it is enough to monitor the error only after the ions moved several MD steps, the total time consumed by evaluating the spilling factor can become insignificantly compared to the total time. So in long molecular dynamics calculations, we recommend using at least {{TAG\|ML_ESTBLOCK\|20-100}}.

	The [[Machine learning force field: Theory#Spilling factor\|spilling factor]] measures the similarity of the local environment of each atom in the current structure to that of the local reference configurations of the force field. The values of the spilling factor are in the range <math>[0,1]</math>. lf the atomic environment is "properly" represented by the local reference configurations the spilling factor approaches 0. Vice versa the spilling factor approaches quickly 1, meaning that the force field is probably extrapolating. Molecular dynamics trajectories where the spilling factor is most of the time 1 can still lead to good results, but the calculations should be cautiously used.		The [[Machine learning force field: Theory#Spilling factor\|spilling factor]] measures the similarity of the local environment of each atom in the current structure to that of the local reference configurations of the force field. The values of the spilling factor are in the range <math>[0,1]</math>. lf the atomic environment is "properly" represented by the local reference configurations the spilling factor approaches 0. Vice versa the spilling factor approaches quickly 1, meaning that the force field is probably extrapolating. Molecular dynamics trajectories where the spilling factor is most of the time 1 can still lead to good results, but the calculations should be cautiously used. <!--You can systematically improve the force field by extracting the structures where the spilling factor is larger than 1 and using [[INTERACTIVE \| interactive mode]] to continue [[Construction:Using metadynamics to train a machine-learned force field \| training the force field specifically on these structures]].-->

	Besides being able to monitor the accuracy during the production runs one can also use the spilling factor to assess the accuracy of the force field for given test sets (i.e. structures chosen from ensembles at different temperatures) like in the traditional way, where test structures are picked out and ab initio calculations have to be carried out for each structure. Using the spilling factor the error is directly assessed without the need for ab initio calculations making the procedure orders of magnitude faster and easier to handle (no evaluation script needed). Nevertheless, if one wants to measure the true error on a test set we have described how to [[Best practices for machine-learned force fields#Test errors\|below]].		Besides being able to monitor the accuracy during the production runs one can also use the spilling factor to assess the accuracy of the force field for given test sets (i.e. structures chosen from ensembles at different temperatures) like in the traditional way, where test structures are picked out and ab initio calculations have to be carried out for each structure. Using the spilling factor the error is directly assessed without the need for ab initio calculations making the procedure orders of magnitude faster and easier to handle (no evaluation script needed). Nevertheless, if one wants to measure the true error on a test set we have described how to [[Best practices for machine-learned force fields#Test errors\|below]].

← Older revision		Revision as of 11:35, 24 October 2025
Line 37:		Line 37:
	The main disadvantage is decreased computational efficiency. The computational cost scales quadratically with the number of species. Using the [[Best practices for machine-learned force fields#Descriptor reduction in production runs\|reduced descriptor]] ({{TAGDEF\|ML_DESC_TYPE\|1}}) can reduce this to linear scaling for major parts of the code, but even perfect linear scaling introduces noticeable overhead when increasing the number of species.		The main disadvantage is decreased computational efficiency. The computational cost scales quadratically with the number of species. Using the [[Best practices for machine-learned force fields#Descriptor reduction in production runs\|reduced descriptor]] ({{TAGDEF\|ML_DESC_TYPE\|1}}) can reduce this to linear scaling for major parts of the code, but even perfect linear scaling introduces noticeable overhead when increasing the number of species.
	{{NB\|warning\|It is not possible to give the same name to different groups of atoms in the {{FILE\|POSCAR}} file and the names are restricted to two characters.}}		{{NB\|warning\|It is not possible to give the same name to different groups of atoms in the {{FILE\|POSCAR}} file and the names are restricted to two characters.}}
	~~'''~~ Ab-initio calculation setup ~~'''~~		==== Ab-initio calculation setup ====

	The training mode requires {{VASP}} to perform ab-initio calculations, so the first step is to set up the [[:Category:Electronic minimization\|electronic minimization]] scheme.		The training mode requires {{VASP}} to perform ab-initio calculations, so the first step is to set up the [[:Category:Electronic minimization\|electronic minimization]] scheme.
Line 48:		Line 48:
	* For simulations without a fixed grid (NpT), the cutoff for plane waves {{TAG\|ENCUT}} must be set at least 30 percent higher than for fixed volume calculations. Also, it is good to restart frequently ({{TAG\|ML_MODE}}=TRAIN with existing {{FILE\|ML_AB}} file in working directory) to reinitialize the [[Projector-augmented-wave formalism\|PAW]] basis of KS orbitals and avoid [[Energy vs volume Volume relaxations and Pulay stress#Pulay stress\|Pulay stress]].		* For simulations without a fixed grid (NpT), the cutoff for plane waves {{TAG\|ENCUT}} must be set at least 30 percent higher than for fixed volume calculations. Also, it is good to restart frequently ({{TAG\|ML_MODE}}=TRAIN with existing {{FILE\|ML_AB}} file in working directory) to reinitialize the [[Projector-augmented-wave formalism\|PAW]] basis of KS orbitals and avoid [[Energy vs volume Volume relaxations and Pulay stress#Pulay stress\|Pulay stress]].
	{{NB\|warning\|It is very important to '''not''' change the ab-initio settings in the {{FILE\|INCAR}} file between training from scratch and continuing training. Likewise, the {{FILE\|POTCAR}} file is '''not''' allowed to be changed when resuming training.}}		{{NB\|warning\|It is very important to '''not''' change the ab-initio settings in the {{FILE\|INCAR}} file between training from scratch and continuing training. Likewise, the {{FILE\|POTCAR}} file is '''not''' allowed to be changed when resuming training.}}
	~~'''~~ Molecular dynamics setup ~~'''~~		==== Molecular dynamics setup ====

	After the forces are obtained from electronic minimization by the [[Hellmann-Feynman forces\|Hellmann-Feynman Theorem]], {{VASP}} must propagate the ions to obtain a new configuration in phase space. For the molecular dynamics part, familiarity with setting up {{TAG\|molecular dynamics}} runs is beneficial. In addition, we recommend the following settings in the molecular dynamics part:		After the forces are obtained from electronic minimization by the [[Hellmann-Feynman forces\|Hellmann-Feynman Theorem]], {{VASP}} must propagate the ions to obtain a new configuration in phase space. For the molecular dynamics part, familiarity with setting up {{TAG\|molecular dynamics}} runs is beneficial. In addition, we recommend the following settings in the molecular dynamics part: