Python Toolkit for Uncertainty Quantification (PyTUQ)

Khachik Sargsyan, Bert Debusschere, Emilie Grace Baillo

Overview

The Python Toolkit for Uncertainty Quantification (PyTUQ) is a Python-only collection of libraries and tools designed for quantifying uncertainty in computational models. PyTUQ offers a range of UQ functionalities, including Bayesian inference and linear regression methods, polynomial chaos expansions, and global sensitivity analysis methods. PyTUQ features advanced techniques for dimensionality reduction, such as SVD and Karhunen-Loeve expansions, along with various MCMC methods for calibration and inference. The toolkit also includes robust classes for multivariate random variables and integration techniques, making it a versatile resource for researchers and practitioners seeking to quantify uncertainty in their numerical predictions. To explore the PyTUQ documentation and learn more, visit our website here.

Dependencies

PyTUQ requires:

• numpy
• scipy
• matplotlib

Optional dependencies include:

• pytorch (NN surrogates)
• QUiNN (Quantification of Uncertainties in Neural Networks)
• pyswarms (Particle Swarm Optimization)

Installation

Basic Installation

Install PyTUQ with core dependencies:

$ pip install pytuq

Optional Features

Install additional features as needed:

• Neural network surrogates

$ pip install pytuq[nn]

• Particle swarm optimization

$ pip install pytuq[optim]

• All optional features

pip install pytuq[all]

• Development tools

pip install pytuq[dev]

Installation from Source

1. To install PyTUQ from source, start up a Python virtual environment and clone the repository:

    $ source <PYTHON_VENV_DIR>/bin/activate
    $ git clone git@github.com:sandialabs/pytuq.git
    $ cd pytuq

2. (Optional) To take advantage of the neural network surrogate model capabilites in PyTUQ, use the following command to install PyTUQ's optional dependencies:

    $ pip install -r requirements.txt

3. Install primary PyTUQ dependencies and PyTUQ:

    $ pip install .

License

Distributed under BSD 3-Clause License. See LICENSE.txt for more information.

Acknowledgements

This work is supported by the Scientific Discovery through Advanced Computing (SciDAC) Program under the Office of Science at the U.S. Department of Energy.

Sandia National Laboratories is a multimission laboratory managed and operated by National Technology & Engineering Solutions of Sandia, LLC, a wholly owned subsidiary of Honeywell International Inc., for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-NA0003525.