Train a Stillinger-Weber potential#

In this tutorial, we train a Stillinger-Weber (SW) potential for silicon that is archived on OpenKIM.

Before getting started to train the SW model, let’s first install the SW model:

$ kim-api-collections-management install user SW_StillingerWeber_1985_Si__MO_405512056662_006

See also

This installs the model and its driver into the User Collection. See Install a model for more information about installing KIM models.

We are going to create potentials for diamond silicon, and fit the potentials to a training set of energies and forces consisting of compressed and stretched diamond silicon structures, as well as configurations drawn from molecular dynamics trajectories at different temperatures. Download the training set Si_training_set.tar.gz. (It will be automatically downloaded if not present.) The data is stored in # extended xyz format, and see Dataset for more information of this format.


The Si_training_set is just a toy data set for the purpose to demonstrate how to use KLIFF to train potentials. It should not be used to train any potential for real simulations.

Let’s first import the modules that will be used in this example.

from kliff.calculators import Calculator
from kliff.dataset import Dataset
from kliff.dataset.weight import Weight
from kliff.loss import Loss
from kliff.models import KIMModel
from kliff.utils import download_dataset


We first create a KIM model for the SW potential, and print out all the available parameters that can be optimized (we call this model parameters).

model = KIMModel(model_name="SW_StillingerWeber_1985_Si__MO_405512056662_006")


# Available parameters to optimize.
# Parameters in `original` space.
# Model: SW_StillingerWeber_1985_Si__MO_405512056662_006

name: A
value: [15.28484792]
size: 1

name: B
value: [0.60222456]
size: 1

name: p
value: [4.]
size: 1

name: q
value: [0.]
size: 1

name: sigma
value: [2.0951]
size: 1

name: gamma
value: [2.51412]
size: 1

name: cutoff
value: [3.77118]
size: 1

name: lambda
value: [45.5322]
size: 1

name: costheta0
value: [-0.33333333]
size: 1

'#================================================================================\n# Available parameters to optimize.\n# Parameters in `original` space.\n# Model: SW_StillingerWeber_1985_Si__MO_405512056662_006\n#================================================================================\n\nname: A\nvalue: [15.28484792]\nsize: 1\n\nname: B\nvalue: [0.60222456]\nsize: 1\n\nname: p\nvalue: [4.]\nsize: 1\n\nname: q\nvalue: [0.]\nsize: 1\n\nname: sigma\nvalue: [2.0951]\nsize: 1\n\nname: gamma\nvalue: [2.51412]\nsize: 1\n\nname: cutoff\nvalue: [3.77118]\nsize: 1\n\nname: lambda\nvalue: [45.5322]\nsize: 1\n\nname: costheta0\nvalue: [-0.33333333]\nsize: 1\n\n'

The output is generated by the last line, and it tells us the name, value, size, data type and a description of each parameter.


You can provide a path argument to the method echo_model_params(path) to write the available parameters information to a file indicated by path.


The available parameters information can also by obtained using the kliff Command Line Tool: $ kliff model --echo-params SW_StillingerWeber_1985_Si__MO_405512056662_006

Now that we know what parameters are available for fitting, we can optimize all or a subset of them to reproduce the training set.

    A=[[5.0, 1.0, 20]], B=[["default"]], sigma=[[2.0951, "fix"]], gamma=[[1.5]]


# Model parameters that are optimized.
# Note that the parameters are in the transformed space if
# `params_transform` is provided when instantiating the model.

A 1
  5.0000000000000000e+00   1.0000000000000000e+00   2.0000000000000000e+01

B 1

sigma 1
  2.0951000000000000e+00 fix

gamma 1

'#================================================================================\n# Model parameters that are optimized.\n# Note that the parameters are in the transformed space if \n# `params_transform` is provided when instantiating the model.\n#================================================================================\n\nA 1\n  5.0000000000000000e+00   1.0000000000000000e+00   2.0000000000000000e+01 \n\nB 1\n  6.0222455840000000e-01 \n\nsigma 1\n  2.0951000000000000e+00 fix \n\ngamma 1\n  1.5000000000000000e+00 \n\n'

Here, we tell KLIFF to fit four parameters B, gamma, sigma, and A of the SW model. The information for each fitting parameter should be provided as a list of list, where the size of the outer list should be equal to the size of the parameter given by model.echo_model_params(). For each inner list, you can provide either one, two, or three items.

  • One item. You can use a numerical value (e.g. gamma) to provide an initial guess of the parameter. Alternatively, the string 'default' can be provided to use the default value in the model (e.g. B).

  • Two items. The first item should be a numerical value and the second item should be the string 'fix' (e.g. sigma), which tells KLIFF to use the value for the parameter, but do not optimize it.

  • Three items. The first item can be a numerical value or the string 'default', having the same meanings as the one item case. In the second and third items, you can list the lower and upper bounds for the parameters, respectively. A bound could be provided as a numerical values or None. The latter indicates no bound is applied.

The call of model.echo_opt_params() prints out the fitting parameters that we require KLIFF to optimize. The number 1 after the name of each parameter indicates the size of the parameter.


The parameters that are not included as a fitting parameter are fixed to the default values in the model during the optimization.

Training set#

KLIFF has a Dataset to deal with the training data (and possibly test data). Additionally, we define the energy_weight and forces_weight corresponding to each configuration using Weight. In this example, we set energy_weight to 1.0 and forces_weight to 0.1. For the silicon training set, we can read and process the files by:

dataset_path = download_dataset(dataset_name="Si_training_set")
weight = Weight(energy_weight=1.0, forces_weight=0.1)
tset = Dataset(dataset_path, weight)
configs = tset.get_configs()


2022-04-28 11:03:34.898 | INFO     | kliff.dataset.dataset:_read:397 - 1000 configurations read from /Users/mjwen/Applications/kliff/examples/Si_training_set

The configs in the last line is a list of Configuration. Each configuration is an internal representation of a processed extended xyz file, hosting the species, coordinates, energy, forces, and other related information of a system of atoms.


Calculator is the central agent that exchanges information and orchestrate the operation of the fitting process. It calls the model to compute the energy and forces and provide this information to the Loss function (discussed below) to compute the loss. It also grabs the parameters from the optimizer and update the parameters stored in the model so that the up-to-date parameters are used the next time the model is evaluated to compute the energy and forces. The calculator can be created by:

calc = Calculator(model)
_ = calc.create(configs)


2022-04-28 11:03:38.460 | INFO     | kliff.calculators.calculator:create:107 - Create calculator for 1000 configurations.

where calc.create(configs) does some initializations for each configuration in the training set, such as creating the neighbor list.

Loss function#

KLIFF uses a loss function to quantify the difference between the training set data and potential predictions and uses minimization algorithms to reduce the loss as much as possible. KLIFF provides a large number of minimization algorithms by interacting with SciPy. For physics-motivated potentials, any algorithm listed on scipy.optimize.minimize and scipy.optimize.least_squares can be used. In the following code snippet, we create a loss of energy and forces and use 2 processors to calculate the loss. The L-BFGS-B minimization algorithm is applied to minimize the loss, and the minimization is allowed to run for a max number of 100 iterations.

steps = 100
loss = Loss(calc, nprocs=2)
loss.minimize(method="L-BFGS-B", options={"disp": True, "maxiter": steps})


2022-04-28 11:03:38.462 | INFO     | kliff.loss:minimize:290 - Start minimization using method: L-BFGS-B.
2022-04-28 11:03:38.462 | INFO     | kliff.loss:_scipy_optimize:406 - Running in multiprocessing mode with 2 processes.
2022-04-28 11:05:36.267 | INFO     | kliff.loss:minimize:292 - Finish minimization using method: L-BFGS-B.

      fun: 0.6940780132834561
 hess_inv: <3x3 LbfgsInvHessProduct with dtype=float64>
      jac: array([-3.65263345e-06,  1.88971060e-04,  3.90798507e-06])
     nfev: 180
      nit: 36
     njev: 45
   status: 0
  success: True
        x: array([14.93863664,  0.58740288,  2.20146242])

The minimization stops after running for 27 steps. After the minimization, we’d better save the model, which can be loaded later for the purpose to do a retraining or evaluations. If satisfied with the fitted model, you can also write it as a KIM model that can be used with LAMMPS, GULP, ASE, etc. via the kim-api.

# model.load("kliff_model.yaml")


# Model parameters that are optimized.
# Note that the parameters are in the transformed space if
# `params_transform` is provided when instantiating the model.

A 1
  1.4938636641779590e+01   1.0000000000000000e+00   2.0000000000000000e+01

B 1

sigma 1
  2.0951000000000000e+00 fix

gamma 1

2022-04-28 11:05:36.291 | INFO     | - KLIFF trained model write to `/Users/mjwen/Applications/kliff/examples/SW_StillingerWeber_1985_Si__MO_405512056662_006_kliff_trained`

The first line of the above code generates the output. A comparison with the original parameters before carrying out the minimization shows that we recover the original parameters quite reasonably. The second line saves the fitted model to a file named kliff_model.pkl on the disk, and the third line writes out a KIM potential named SW_StillingerWeber_1985_Si__MO_405512056662_006_kliff_trained.

See also

For information about how to load a saved model, see Frequently Used Modules.

Total running time of the script: ( 2 minutes 4.046 seconds)

Gallery generated by Sphinx-Gallery