Best practices for machine-learned force fields

Using the machine-learning–force-fields method, VASP can construct force fields based on ab-initio simulations. There are many aspects to carefully consider while constructing, testing, retraining, and applying a force field. Here, we list some best practices but bare in mind that the list is incomplete and the method has not been applied to a large number of systems. Therefore, we recommend using the usual rigorous checking that is necessary for any research project.

Training

We strongly advise to follow this guideline in on-the-fly learning:

• It is generally possible to train force fields on a smaller unit cell and then apply it to a larger system. Be careful to choose the structure large enough, so that the phonons or collective vibrations "fit" into the supercell.
• It is important to learn the correct forces. This requires converged electronic calculations with respect to electronic parameters, i.e., the number of k points, the plane-wave cutoff (ENCUT), electronic minimization algorithm, etc.
• Switch off the symmetry as in molecular-dynamics runs (ISYM=0).
• Lower the integration step (POTIM) if the system contains light elements or increase the mass of the light elements (POMASS) in the INCAR or POTCAR file.
• If possible, heat up the system gradually, i.e., use temperature ramping from a low temperature to a temperature approximately 30% higher than the desired application temperature. The determination of the threshold of the maximum error in forces for Bayesian prediction (ML_CTIFOR) is done on the basis of previous Bayesian error prediction of forces. At the beginning of the on-the-fly learning the force field is very inaccurate and also the prediction of Bayesian errors is inaccurate. In many cases this can lead to an overestimation of ML_CTIFOR. If this happens no further learning will be done since each forthcoming structure has a predicted error that is smaller than the threshold, although not enough training structures and local reference configurations have been collected and the force field has a bad accuracy. The heating of the systems greatly helps to avoid these "unlucky situations", because the error of the system increases with the temperature and the threshold for will be easily exceeded, leading to further learning and improvement of the force field. Since the error of the system decreases with decreasing the temperature, the threshold would be always stuck at the high temperature errors in cooling runs. Hence we don't advise to use cooling runs for on-the-fly learning. Finally, when using temperature ramps, always be very careful to work in a temperature region where no phase transitions are induced.
• If possible, prefer molecular-dynamics training runs in the NpT ensemble, the additional cell fluctuations improve the robustness of the resulting force field. However, for liquids allow only scaling of the supercell (no change of angles, no individual lattice vector scaling) because otherwise the cell may collapse. This can be achieved with the ICONST file, see 3) and 4) here.
• For simulations without fixed lattice, the plane-wave cutoff ENCUT needs to be set at least 30 percent higher than for fixed volume calculations. Also it is goot to restart often (ML_ISTART=1) to reinitialize the PAW basis of the KS orbitals and avoid Pulay stress.
• When the system contains different components, train them separately first. For instance, when the system has a surface of a crystal and a molecule binding on that surface. First, train the bulk crystal, then the surface, next the isolated molecule, and finally the entire system. This way a significant amount of ab-initio calculation can be saved in the combined system which is anyway the computationally most expensive.

Monitoring

The monitoring of your learning can be divided into two parts:

• Molecular-dynamics/ensemble related quantities:
  • Monitor your structure visually. This means look at the CONTCAR or XDATCAR files with structure/trajectory viewers. Many times the fact that something goes wrong can be immediately caught when the system undergoes some unwanted or unphysical deformations.
  • Volume and lattice parameters in the OUTCAR and CONTCAR files. It is important to confirm that the average volume stays in the desired region. Strong change of the average volume over time in constant temperatur and pressure runs indicates phase transitions or non-properly equilibrated systems.
  • Temperature and pressure in the OUTCAR and OSZICAR files. Strong deviations of temperature and pressure with respect to desired ones at the beginning of the calculation indicates non-properly equilibrated systems.
  • Pair-correlation functions (PCDAT).
• Machine learning specific quantities in the ML_LOGFILE file:
  • ERR: Root mean square error of predicted energy, forces and stress with respect to ab-initio data for all training structures up to the current molecular-dynamics step.
  • BEEF: Estimated Bayesian error of energy, forces and stress (columns 3-5). Current threshold for the maximum Bayesian error of forces ML_CTIFOR on column 6.
  • THRUPD: Update of ML_CTIFOR.
  • THRHIST: History of Bayesian errors used for ML_CTIFOR.

Tuning on-the-fly parameters

In case too many or too few training structures are selected there are some on-the-fly parameters which can be tuned:

• ML_CTIFOR: In a continuation run it can be set to the last value of ML_CTIFOR of the previous run. This way unneccesary sampling at the beginning of the calculation can be skipped. However, when going from one structure to the other, this tag should be very carefully set. ML_CTIFOR is species and system dependent. Low symmetry structures, for example liquids, have usually a much higher error than high symmetry solids for the same compound. If a liquid is learned first and the last ML_CTIFOR from the liquid is used for the corresponding solid, this ML_CTIFOR is way too large for the solid and all predicted errors will be below the threshold. Hence no learning will be done on the solid. In this case it is better to start with the default value of ML_CTIFOR.
• ML_CX:
• ML_HIS
• ML_SCLC_CTIFOR

If not enough or too many local reference structures are extracted from the training data there are additional tags for tuning:

• ML_EPS_LOW

Testing

• Set up an independent test set of random configurations. Then, check the average errors comparing forces, stresses and energy differences of two structures based on DFT and predicted by machine-learned force fields.

• If you have both, ab-initio reference data and a calculation using force fields, check the agreement of some physical properties. For instance, you might check the relaxed lattice parameters, phonons, relative energies of different phases, the elastic constant, the formation of defects etc.

• Plot blocked averages

Application

The following things need to be considered when running only the force field (ML_ISTART=2):

• Set the ab-initio parameters to small values. VASP cannot circumvent the initialization of KS orbitals although they are not used during the molecular-dynamics run with machine learning.

Example

Learning of liquid water in the NpT ensemble.

INCAR:

ENCUT  = 520 #larger cutoff
LASPH  = .True.

IBRION = 0
MDALGO = 3
ISIF   = 3
POTIM  = 1.5
TEBEG  = 200
TEEND  = 500
LANGEVIN_GAMMA = 10.0 10.0
LANGEVIN_GAMMA_L = 3.0
PMASS = 100
PSTRESS = 0.001

POMASS = 8.0 16.0

ML_LMLFF  = .TRUE.
ML_ISTART = 0

ICONST:

LA 1 2 0
LA 1 3 0
LA 2 3 0
LR 1 0
LR 2 0
LR 3 0
S  1  0  0  0  0  0 0
S  0  1  0  0  0  0 0
S  0  0  1  0  0  0 0
S  0  0  0  1 -1  0 0
S  0  0  0  1  0 -1 0
S  0  0  0  0  1 -1 0

Related articles

Machine-learned force fields, Ionic minimization, Molecular dynamics