Skip to content

atomicarchitects/phonax

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

20 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Phonax

Phonax is a JAX framework to use E(3)-equivariant neural network for predicting vibrational and phonon modes of molecules and periodic crystals. Alternatively, Phonax also enables the utilization of the second-order derivative Hessians or the vibrational spectrum as eigenvalues to fine-tune the underlying energy models.

Quick Start

The following steps will guide you through the installation for the phonax library and the tests with tutorial notebooks.

  1. In a conda environment with JAX / CUDA available, clone and download this phonax repository.

    git clone https://github.com/atomicarchitects/phonax.git
    cd phonax
  2. In the downloaded directory run:

    pip install -e .
  3. Install the additional libraries from github repositories:

    pip install -r requirements_extra.txt
  4. Download the pre-trained model weights and the datasets, and uncompress the file at the folder with the tutorial notebooks.

    tar zxvf phonax-download.tar.gz

Tutorials and scripts

  1. Classical mass and spring mechanical model for phonon vibrations Tutorial

    A classical mass and spring model can be used to give intuitions for the vibrational phonon modes. With the simple energy functional form (energy stored in the mechanical springs), this tutorial shows the main architecture and usage of the phonax in deriving the second derivative Hessians. As an extension, we also give a prototype phonax interface with PyTorch, since many neural network models in the literature are written with PyTorch library.

  2. Periodic crystalline solids phonon predictions with pre-trained energy models Tutorial

    To showcase the phonon band predictions for general crystalline solids, we provide several pre-trained energy models (NqeuIP and MACE, with PBE and PBEsol) that cant be used to derive the phonon properties, given a crystal structure file (vasp format).

  3. Train a new energy model for making phonon predictions Tutorial

    For specific material applications, one can train new energy models that are tailored to these materials by training energy and force data generated for the relevant material types. With a converged energy model, one can make better phonon predictions for these applications. To examine the equilibrium condition for the given crystal structure, one can check the atomic forces, which should be vanishing for a meaningful phonon prediction.

  4. IR / Raman optical spectroscopy activities for vibrational states Tutorial

    Besides making vibrational mode energy predictions, phonax can also be used to analyze the vibrational state symmetry properties and their activities under IR / Raman optical spectroscopy probes from the symmetry selection rules.

  5. Hessian training for molecules Tutorial

    In this notebook, we demonstrate how to improve and fine-tune the energy model by augmented training with molecular Hessians or the vibrational spectrum (eigenvalues of the dynamical matrix from hessians)

  6. Hessian training for crystals Tutorial

    Similarily, the Hessian training can also be used for the periodic crystals to improve the energy model and foce predictions.

  7. Gradio phonax prediction interface

    To launch this gradio web interface for using phonax, run the follwing command in the download folder:

    python launch_gradio_service.py

    A default local URL will be generated at http://0.0.0.0:7860 which can be opened in any web browser. In this gradio web interface, one can simply upload a structural file in cif format, and choose the prediction setting (select the fine-tuned model for PBEsol predictions, otherwise PBE based predictions). The phonon band structure and the (projected) phonon density of state will be computed.

Models for generalized energy potential with electric fields and tensorial physical quantity predictions for dipole-dipole interactions that correct long-wavelength phonon modes

Please check out the GitHub repo

References

  1. e3nn: Euclidean Neural Networks

  2. NequIP: E(3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials

  3. MACE: Higher Order Equivariant Message Passing Neural Networks for Fast and Accurate Force Fields, The Design Space of E(3)-Equivariant Atom-Centered Interatomic Potentials

  4. JAX M.D. A Framework for Differentiable Physics

  5. PoSym: A python library to analyze the symmetry of theoretical chemistry objects

Citation

  1. NeurIPS 2023 AI4Mat workshop

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published