CGCNN

Description

The Crystal Graph Convolutional Neural Networks (CGCNN) software package implements a sophisticated AI model capable of predicting material properties based on arbitrary crystal structures. This tool is invaluable for researchers and developers in materials science seeking to understand and forecast the characteristics of new materials.

CGCNN operates by processing crystal structures as graphs, where atoms are nodes and bonds are edges. Convolutional neural networks are then applied to these graphs to learn representations that correlate with material properties. The package offers two primary functionalities: training a CGCNN model with a user-defined dataset and predicting material properties for new crystals using a pre-trained CGCNN model.

To train a custom model, users must first define a dataset comprising CIF files for crystal structures and a corresponding CSV file (id_prop.csv) linking crystal IDs to their target properties. An initialization vector file (atom_init.json) is also required. The training process can be initiated using the `main.py` script, allowing for customization of training, validation, and test data splits. The output includes saved model checkpoints and test results.

For prediction tasks, users need a pre-trained CGCNN model file (e.g., `pre-trained.pth.tar`) and a dataset of crystal structures for which properties are to be predicted. The `predict.py` script takes these inputs and generates a `test_results.csv` file containing the predicted properties for each crystal. This functionality is crucial for rapid screening of potential materials and accelerating the discovery process.

The CGCNN framework is built upon PyTorch and requires several prerequisites, including scikit-learn and pymatgen. Installation is streamlined through conda environments. The project is open-source, released under the MIT License, and is a testament to advancements in applying deep learning to materials science challenges.

CGCNN Highlights

• Predicts material properties from crystal structures

• Trains custom CGCNN models with user datasets

• Utilizes pre-trained CGCNN models for property prediction

• Processes crystal structures as graphs

• Supports both regression and classification tasks

• Outputs trained model checkpoints

• Generates test results with predicted properties

• Requires PyTorch, scikit-learn, and pymatgen

• Compatible with PyTorch v1.0.0+

Getting Started with CGCNN

1. 1

   Set up environment: Install PyTorch, scikit-learn, and pymatgen via conda.

2. 2

   Define dataset: Prepare CIF files for crystal structures and an id_prop.csv file.

3. 3

   Train model: Run `python main.py <root_dir>` to train a custom CGCNN model.

4. 4

   Predict properties: Use `python predict.py <model_path> <root_dir>` with a pre-trained model.

5. 5

   Activate environment: Use `source activate cgcnn` before running scripts.

6. 6

   Deactivate environment: Use `source deactivate` after finishing.

CGCNN's Use Cases

• Material Property Prediction
• Crystal Structure Analysis
• Materials Discovery
• Computational Materials Science
• Predictive Modeling
• Data-Driven Materials Design

FAQ from CGCNN

• What is CGCNN used for?

  CGCNN is used to predict material properties from crystal structures using crystal graph convolutional neural networks.

• What are the prerequisites for using CGCNN?

  CGCNN requires PyTorch, scikit-learn, and pymatgen. It is tested for PyTorch v1.0.0+.

• How do I train a custom CGCNN model?

  You need to define a customized dataset with CIF files and an id_prop.csv, then run `python main.py <root_dir>`.

• Can I predict material properties without training a model?

  Yes, you can use a pre-trained CGCNN model with `python predict.py <model_path> <root_dir>`.

• What kind of data format does CGCNN accept for crystal structures?

  CGCNN accepts crystal structures in CIF file format.

• What is the output of the prediction script?

  The prediction script outputs a `test_results.csv` file containing IDs and predicted properties for each crystal.

• Is CGCNN suitable for both regression and classification tasks?

  Yes, CGCNN supports both regression and classification tasks, which can be specified during training.

• What license is CGCNN released under?

  CGCNN is released under the MIT License.

• Where can I find example datasets for CGCNN?

  Example datasets for regression and classification are available in the repository under `data/sample-regression` and `data/sample-classification`.

• How do I install CGCNN and its dependencies?

  It is recommended to install prerequisites using conda by creating a new environment: `conda create -n cgcnn python=3 scikit-learn pytorch torchvision pymatgen -c pytorch -c conda-forge`.