KIM-based Learning-Integrated Fitting Framework (KLIFF)

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.

Installation

Using conda (recommended)

conda install -c conda-forge kliff

Using pip

pip install kliff

From source

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

Dependencies

• KLIFF requires the kim-api and kimpy packages. If you install using conda as described above, these packages will be installed automatically. Alternatively, they can be installed from source or via pip. See Installation for more information.

• In addition, PyTorch is needed if you want to train machine learning models. See the official PyTorch website for installation instructions.

A quick example to train a neural network potential

from kliff.legacy import nn
from kliff.legacy.calculators import CalculatorTorch
from kliff.legacy.descriptors import SymmetryFunction
from kliff.dataset import Dataset
from kliff.models import NeuralNetwork
from kliff.legacy.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)
model.add_layers(
    # first hidden layer
    nn.Linear(descriptor.get_size(), N1),
    nn.Tanh(),
    # second hidden layer
    nn.Linear(N1, N2),
    nn.Tanh(),
    # 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.from_path(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)
calc.create(configs)

# Define a loss function and train the model by minimizing the loss
loss = Loss(calc)
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 and ASE
model.write_kim_model()

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)
• Support for deploying trained ML models with TorchScript and KIM-API

Citing KLIFF

@Article{wen2022kliff,
  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},
  volume  = {272},
  pages   = {108218},
  year    = {2022},
  doi     = {10.1016/j.cpc.2021.108218},
}