🌐 Crystal Graph Neural Networks 💎

A banchmarking framework for Crystal Graph Neural Networks (CGNN)

built on top of PyTorch Lightning and Deep Graph Library (DGL)

📜 Table of Contents

• 1. Introduction
• 2. Installation
• 3. Usage
  • 1. Benchmarking with the Jarvis-Tools Database
  • 2. Benchmarking with the Matbench Database
• 4. Results and Benchmarks
  • 1. Jarvis-Tools Database
  • 2. Matbench Database
• 5. Contributing
  • Adding a New Model
• 6. License

---

1. Introduction

Welcome to the world of Crystal Graph Neural Networks (CGNN)! This exciting field of research finds its applications in diverse areas, ranging from material science to chemistry. This repository has been created to serve as a comprehensive benchmarking framework tailored for CGNNs. It aims to facilitate the evaluation of CGNN performance, making it easier to assess their capabilities.

This repository have integrated well-known models and databases, such as Jarvis-Tools and Matbench, making use of the power and flexibility of PyTorch Lightning and DGL.

🚀 Features

• Supported Models: (more models will be added)

  • SCHNET
  • CGCNN
  • ALIGNN

• database Integrated:

  • Jarvis-Tools database
  • Matbench

• Performance Evaluation: Post-benchmarking, model performances can be visualized and compared using Tensorboard.

---

2. Installation

Method 1: Reproducible Installation

This method ensures a more reproducible setup by using a pre-defined environment.

git clone https://github.com/hspark1212/crystal-gnn.git
cd crystal-gnn
conda env create -f environment.yml

Method 2: Standard Installation

This method is more flexible without strict dependencies.

Step 1: Make Conda Environment with Python 3.9

conda create -n crystal-gnn python=3.9
conda activate crystal-gnn

Step 2: Clone the Repository & Install Dependencies

git clone https://github.com/hspark1212/crystal-gnn.git
cd crystal-gnn
pip install -r requirements.txt

---

3. Usage

1) Benchmarking with the Jarvis-Tools Database

The script benchmark_jarvis.sh is an example of how to run the benchmarking across all models on the Jarvis-Tools database. Once the training is completed, results are stored in the ./crystal-gnn/logs/ directory. Visualization of these results is made possible through Tensorboard.

Example 1: To benchmark the CGCNN model on the formation_energy_per_atom task, run the following command:

python run.py with cgcnn exp_name=cgcnn_formation_energy_peratom database_name=dft_3d_2021 target=formation_energy_peratom

This command will download the Jarvis-Tools database in ./crystal-gnn/data and train the CGCNN model on the formation_energy_per_atom task. The trained model will be saved in ./logs/cgcnn_formation_energy_peratom/.

Example 2: To benchmark the ALIGNN model on the band gap task and set hidden_dim to 32, run the following command:

python run.py with alignn exp_name=alignn_optb88vdw_bandgap database_name=dft_3d_2021 target=optb88vdw_bandgap hidden_dim=32

Just as you can adjust the hidden_dim parameter, various parameters related to targets, training, logs, and individual models can be modified in the configs.py file.

Viewing Results with Tensorboard:

After completing the benchmarks, launch Tensorboard with the following command to visualize the results:

tensorboard --logdir=./crystal-gnn/logs/ --bind_all

By entering the provided URL into your web browser (usually something like http://localhost:6006/), you can access Tensorboard's graphical interface and review the benchmark outcomes.

2) Benchmarking with the Matbench Database

The script eval_matbench.py is to benchmark using the Matbench database, which includes 13 different tasks. Among these tasks, only 7 tasks involve utilizing structural information as inputs. Consequently, this script is designed to perform benchmarking on 6 regression tasks and 1 classification task, listed as follows:

matbench_log_gvrh
matbench_log_kvrh
matbench_mp_e_form
matbench_mp_gap
matbench_mp_is_metal
matbench_perovskites
matbench_phonons

Example 1: To run benchmarks for the CGCNN model on the listed 7 tasks, run the following command:

python eval_matbench.py with model_name=cgcnn target=all

Example 2: To run benchmarks for the CGCNN model on the specific task matbench_mp_gap, run the following command:

python eval_matbench.py with model_name=cgcnn target=matbench_mp_gap

After training, the model along with the corresponding log files will be stored at ./logs/{model_name}_matbench_{task_name}. Furthermore, the JSON files, intended for submission to the Matbench leaderboard, will be stored at ./logs/{model_name}_{task_name}/results_{model_name}_{task_name}.json.gz.

---

4. Results and Benchmarks

The benchmarks are implemented with the following dependencies:

• OS: Ubuntu 20.04
• GPU: NVIDIA GeForce RTX 2080 Ti
• CUDA: 11.7
• cuDNN: 8.9.3
• python: 3.9
• Dependencies:

  dgl==1.1.2 (cu117)
  torch==2.0.1
  pytorch-lightning==2.0.6
  

For reproducibility, our benchmarks have been verified using pytorch-lightning configurations (seed_everything=0 and deterministic=True). It will guarantee the same results in the same machine and environment.

1) Jarvis-Tools Database

Task	SCHNET	CGCNN	ALIGNN
formation_energy_per_atom	0.057 (0.98)	0.039 (0.99)	0.034 (0.99)
Band gap	0.194 (0.85)	0.147 (0.89)	0.133 (0.89)
energy above hull	0.162 (0.94)	0.106 (0.98)	0.066 (0.99)
bulk modulus	15.159 (0.83)	13.271 (0.85)	14.043 (0.85)
shear modulus	11.710 (0.70)	10.960 (0.74)	10.497 (0.73)

This results are stored in results/benchmark_jarvis, which can be visualized with the following command tensorboard --logdir=./crystal-gnn/results/benchmark_jarvis --bind_all.

2) Matbench Database

Task	SCHNET	CGCNN	ALIGNN (will be updated soon)
matbench_log_gvrh	0.097 (0.84)	0.090 (0.86)
matbench_log_kvrh	0.075 (0.87)	0.068 (0.88)
matbench_mp_e_form	0.039 (-56.24)	0.027 (-37.33)
matbench_mp_gap	0.287 (0.57)	0.206 (0.86)
matbench_perovskites	0.051 (0.98)	0.039 (0.99)
matbench_phonons	87.952 (0.81)	65.544 (0.91)
matbench_mp_is_metal	[0.898]	[0.899]

• The results are computed using the average of five folds.

Note: Results are displayed as MAE (R2) for regression tasks and [Accuracy] for classification tasks.

---

5. Contributing

Other databases and models are welcome to be added to the benchmarking. If you would like to contribute, please open an issue or submit a pull request.

🤗 Adding a New Model

The following steps are required to add a new model to the benchmarking:

Step 1: Create a new model class in the models directory that inherits from the BaseModule class. The BaseModule class is a wrapper class that inherits from pytorch-lightning's LightningModule class and provides a convenient interface for training and evaluation. The BaseModule class also provides a forward method that takes in a dgl.DGLGraph object and returns a prediction. The forward method is used for inference and evaluation.

from crystal_gnn.models.base_module import BaseModule

class MyModel(BaseModule):
    def __init__(self, _config: Dict[str, Any]):
        super().__init__(_config)
        # Initialize your model here

    def forward(self, batch: Dict[str, Any]) -> torch.Tensor:
        # Perform inference and return a prediction
        graph = batch['graph']
        line_graph = batch['line_graph'] # optional
        # Message passing ...

The batch includes graph and line graph representations of the input data. The graph is a dgl.DGLGraph object that contains the following node and edge features:

• g.ndata['atomic_number']: Atomic number of each node (torch.Tensor, shape = (N, 1))
• g.ndata['coord']: Atomic coordinates of each node (torch.Tensor, shape = (N, 3))
• g.edata['distance']: Distance between two nodes (torch.Tensor, shape = (E, 1))
• g.edata['angle']: Angle between three nodes (torch.Tensor, shape = (E, 1)) (If line_graph is provided)

For simplicity, the hyperparmeters of models are defined as follows:

• model_name: Name of the model (str, "schnet", "cgcnn", "alignn")
• num_conv: Number of convolution layers (int, default = 4)
• hidden_dim: Hidden dimension of the model (int, default = 256)
• rbf_distance_dim: RDF expansion dimension for edge distance (int, default = 80)
• rbf_triplet_dim: RDF expansion dimension for triplet angle (int, default = 40)
• batch_norm: Whether to use batch normalization (bool, default = True)
• residual: Whether to use residual connections (bool, default = True)
• dropout: Dropout rate (float, default = 0.0)

The hyperparmeters can be added and modified in the configs.py file.

Step 2: Add the model to the _model dictionary in the crystal_gnn/models/init.py file.

from crystal_gnn.models.cgcnn import CGCNN
from crystal_gnn.models.alignn import ALIGNN
from crystal_gnn.models.schnet import SCHNET
from crystal_gnn.models.my_model import MyModel

_models = {
    "cgcnn": CGCNN,
    "alignn": ALIGNN,
    "schnet": SCHNET,
    "my_model": MyModel
}

Step 3: run the benchmarking script with the new model. For example, to run the benchmarking with the my_model model, run the following command:

python run.py with model_name=my_model exp_name=my_model_formation_energy_peratom database_name=dft_3d_2021 target=formation_energy_peratom

---

6. License

This repository is licensed under the MIT License. See LICENSE for more details.

---