KLIFF: KIM-based Learning-Integrated Fitting Framework

fitting, interatomic-potentials, kim, machine-learning
pip install kliff==0.3.1


KIM-based Learning-Integrated Fitting Framework (KLIFF)

Build Status Python package Documentation Status Anaconda-Server Badge PyPI

Documentation at: https://kliff.readthedocs.io

KLIFF is an interatomic potential fitting package that can be used to fit physics-motivated (PM) potentials, as well as machine learning potentials such as the neural network (NN) models.


Using conda

conda intall -c conda-forge kliff

Using pip

pip install kliff

From source

git clone https://github.com/openkim/kliff
pip install ./kliff

To train a KIM model, kim-api and kimpy are needed; to train a machine learning model, PyTorch is needed. For more information on installing these packages, see Installation.

A quick example to train a neural network potential

from kliff import nn
from kliff.calculators import CalculatorTorch
from kliff.descriptors import SymmetryFunction
from kliff.dataset import Dataset
from kliff.models import NeuralNetwork
from kliff.loss import Loss
from kliff.utils import download_dataset

# Descriptor to featurize atomic configurations
descriptor = SymmetryFunction(
    cut_name="cos", cut_dists={"Si-Si": 5.0}, hyperparams="set51", normalize=True

# Fully-connected neural network model with 2 hidden layers, each with 10 units
N1 = 10
N2 = 10
model = NeuralNetwork(descriptor)
    # first hidden layer
    nn.Linear(descriptor.get_size(), N1),
    # second hidden layer
    nn.Linear(N1, N2),
    # output layer
    nn.Linear(N2, 1),

# Training set (dataset will be downloaded from:
# https://github.com/openkim/kliff/blob/master/examples/Si_training_set.tar.gz)
dataset_path = download_dataset(dataset_name="Si_training_set")
dataset_path = dataset_path.joinpath("varying_alat")
train_set = Dataset(dataset_path)
configs = train_set.get_configs()

# Set up calculator to compute energy and forces for atomic configurations in the
# training set using the neural network model
calc = CalculatorTorch(model, gpu=False)

# Define a loss function and train the model by minimizing the loss
loss = Loss(calc, residual_data={"forces_weight": 0.3})
result = loss.minimize(method="Adam", num_epochs=10, batch_size=100, lr=0.001)

# Write trained model as a KIM model to be used in other codes such as LAMMPS ans ASE

Detailed explanation and more tutorial examples can be found in the documentation.

Why you want to use KLIFF (or not use it)

  • Interacting seamlessly with KIM, the fitted model can be readily used in simulation codes such as LAMMPS and ASE via the KIM API
  • Creating mixed PM and NN models
  • High level API, fitting with a few lines of codes
  • Low level API for creating complex NN models
  • Parallel execution
  • PyTorch backend for NN (include GPU training)


  title   = {{KLIFF}: A framework to develop physics-based and machine learning interatomic potentials},
  author  = {Mingjian Wen and Yaser Afshar and Ryan S. Elliott and Ellad B. Tadmor},
  journal = {Computer Physics Communications},
  pages   =  {108218},
  year    = {2021},
  doi     = {10.1016/j.cpc.2021.108218},