https://github.com/google-deepmind/flows_for_atomic_solids
Science Score: 23.0%
This score indicates how likely this project is to be science-related based on various indicators:
-
○CITATION.cff file
-
○codemeta.json file
-
○.zenodo.json file
-
✓DOI references
Found 1 DOI reference(s) in README -
✓Academic publication links
Links to: iop.org -
○Academic email domains
-
○Institutional organization owner
-
○JOSS paper metadata
-
○Scientific vocabulary similarity
Low similarity (13.5%) to scientific vocabulary
Repository
Basic Info
- Host: GitHub
- Owner: google-deepmind
- License: apache-2.0
- Language: Python
- Default Branch: main
- Size: 78.1 KB
Statistics
- Stars: 44
- Watchers: 6
- Forks: 2
- Open Issues: 0
- Releases: 0
Metadata Files
README.md
Flows for atomic solids
The code in this repository can be used to train normalizing flow models to generate samples of atomic solids, as described in our paper Normalizing flows for atomic solids. It also contains a Colab notebook that loads parameters of already trained models and samples from them, plotting observables similar to the figures in the paper.
Installation and usage
Structure of the code
The code is organized in the following folders:
colab: contains a Colab notebook to explore the samples from pre-trained models.experiments: configuration files for Lennard-Jones and monatomic water experiments, and the script to run training on them.models: modules to build normalizing flow models.systems: definitions of the Lennard-Jones and monatomic water potentials used to train the models.tutorial: contains a Colab notebook to train a small model from scratch on the 8-particle monatomic water system with some code to be implemented.utils: utilities for building lattices and computing observables from the model samples.
Training a model
Python version >= 3.7 is required to install and run the code.
To train one of the normalizing flows described in the paper, first clone the deepmind-research repository in a folder of your choice:
shell
git clone https://github.com/deepmind/flows_for_atomic_solids.git
Set up a Python virtual environment with the required dependencies by running
the run.sh script. This will also test-run the training script to
make sure the installation succeeded.
shell
source ./flows_for_atomic_solids/run.sh
Then run the experiments/train.py script, selecting one of the
pre-configured systems:
shell
python -m flows_for_atomic_solids.experiments.train --system='lj_32'
Please note that a GPU is necessary to train the larger models.
Exploring a pre-trained model
The Colab notebook colab/explore_trained_models.ipynb can be used
to access parameters of a set of models that have been trained as described
in the paper. The Colab will load the model and reproduce the energy,
radial distribution and work figures, as well as compute a free-energy
estimate. A Colab runtime with GPU or TPU accelerator is recommended.
Disclaimer
This is not an official Google product.
Owner
- Name: Google DeepMind
- Login: google-deepmind
- Kind: organization
- Website: https://www.deepmind.com/
- Repositories: 245
- Profile: https://github.com/google-deepmind
GitHub Events
Total
- Watch event: 6
Last Year
- Watch event: 6
Dependencies
- absl-py >=0.13.0
- chex >=0.1.1
- distrax >=0.1.1
- dm-haiku >=0.0.6
- dm-tree >=0.1.6
- jax >=0.3.2
- jaxlib >=0.3.2
- ml_collections >=0.1.1
- numpy >=1.21.5
- optax >=0.1.0
- tensorflow-probability >=0.15.0
- absl-py ==1.3.0
- chex ==0.1.5
- cloudpickle ==2.2.0
- contextlib2 ==21.6.0
- decorator ==5.1.1
- distrax ==0.1.2
- dm-haiku ==0.0.8
- dm-tree ==0.1.7
- etils ==0.8.0
- gast ==0.5.3
- importlib-resources ==5.10.0
- jax ==0.3.23
- jaxlib ==0.3.22
- jmp ==0.0.2
- ml-collections ==0.1.1
- numpy ==1.23.4
- opt-einsum ==3.3.0
- optax ==0.1.3
- pyyaml ==6.0
- scipy ==1.9.3
- six ==1.16.0
- tabulate ==0.9.0
- tensorflow-probability ==0.18.0
- toolz ==0.12.0
- typing-extensions ==4.4.0
- zipp ==3.10.0