guaguabujianle/LEIGNN

LEIGNN

Dataset

We use ISO17, H2O, and CH4 as examples to illustrate how to use LEIGNN.

• ISO17 Dataset [1]: Available at ISO17.
• OC20 Dataset [2]: Available at OC20 (Train and Val) and OC20 (Test).
• H2O and CH4 Datasets: Available at H2O and CH4.
• LiPS: Available at LiPS.

Requirements

Required Python packages include:

• ase==3.22.1
• config==0.5.1
• lmdb==1.4.1
• matplotlib==3.7.2
• numpy==1.24.4
• pandas==2.1.3
• pymatgen==2023.5.10
• scikit_learn==1.3.0
• scipy==1.11.4
• torch==1.13.1
• torch_geometric==2.2.0
• torch_scatter==2.1.0
• tqdm==4.66.1

Alternatively, install the environment using the provided YAML file at ./environment/environment.yaml.

Logger

For logging, we recommend using wandb. More details are available at https://wandb.ai/. Training logs and trained models are stored in the ./wandb directory.

Step-by-Step Guide

Data Preprocessing

Download the data from ISO17 + H2O + CH4, OC20 (Train and Val), OC20 (Test) and LiPS. The downloaded data are preprocessed by default. If you wish to preprocess them from scratch, run:

• python preprocess_iso17.py --data_root /path/to/iso17 --num_workers 8 for the ISO17 dataset.
• python preprocess_md.py --data_root /path/to/CH4 --num_workers 8 for the CH4 dataset.
• python preprocess_md.py --data_root /path/to/H2O --num_workers 8 for the H2O dataset.

Replace /path/to/ with your directory paths.

Train the Model

To train LEIGNN, run:

• python train_iso17.py --data_root /path/to/iso17 --num_workers 4 for ISO17.
• python train_oc20.py --data_root /path/to/oc20/200k --data_type 50K --model_type leignn --num_workers 4 --batch_size 32 for OC20-50K and python train_oc20.py --data_root /path/to/oc20/200k --data_type 200K --model_type leignn --num_workers 4 --batch_size 32 for OC20-200K.
• python train_md.py --data_root /path/to/CH4 --systems CH4 --num_workers 4 for CH4.
• python train_md.py --data_root /path/to/H2O --systems H2O --num_workers 4 for H2O.

Replace /path/to/ with your directory paths.

Ablation study

To perform the ablation study, use the following commands:

• python train_oc20.py --data_root /path/to/oc20/200k --data_type 50K --model_type vanilla --num_workers 4 --batch_size 32 for OC20-50K and python train_oc20.py --data_root /path/to/oc20/200k --data_type 200K --model_type leignn --num_workers 4 --batch_size 32 for Vanilla.
• python train_oc20.py --data_root /path/to/oc20/200k --data_type 50K --model_type vanilla_nmu --num_workers 4 --batch_size 32 for OC20-50K and python train_oc20.py --data_root /path/to/oc20/200k --data_type 200K --model_type leignn --num_workers 4 --batch_size 32 for Vanilla + NMU.

To test the case for 200K, simply replace --data_type 50K with --data_type 200K in the commands above.

Test the Model

To test LEIGNN on ISO17, run:

• python test_iso17.py --data_root /path/to/iso17 --model_dir ./wandb/run-20231031_144315-LEIGNN_20231031_144314/ To test LEIGNN on oc20, run:
• python test_oc20.py --data_root /path/to/oc20/200k --data_type 50K --model_dir ./wandb/run-20231031_144315-LEIGNN_20231031_144314/ --model_type leignn --batch_size 32

Replace /path/to/ and ./wandb/run-20231031_144315-LEIGNN_20231031_144314/ with your directory path.

MD Simulations

After training LEIGNN on the LiPS, H2O, and CH4 datasets, run MD simulations with the following commands:

H20

• python simulate_md.py --data_root /path/to/H2O --model_dir ./wandb/run-20231124_233309-LEIGNN_H2O_20231124_233308 Replace /path/to/ and ./wandb/run-20231124_233309-LEIGNN_H2O_20231124_233308 with your directory path.
  Evaluate and visualize MD simulation results using performance_H2O.ipynb

CH4

• python simulate_md.py --data_root /path/to/CH4 --model_dir ./wandb/run-20231124_233309-LEIGNN_CH4_20231124_233308 Evaluate and visualize MD simulation results using performance_CH4.ipynb

LiPS

• python simulate_lips.py --data_root /path/to/lips/20k --model_dir ./wandb/run-20240717_202344-LEIGNN_LiPS_20240717_202343
  Evaluate and visualize MD simulation results using performance_lips.py

Acknowledgements

Some part of code in this project were adapted from OCP and MDsim. We gratefully acknowledge the contributions from these sources.

Reference

[1] Schütt K, Kindermans P J, Sauceda Felix H E, et al. "Schnet: A continuous-filter convolutional neural network for modeling quantum interactions." Advances in Neural Information Processing Systems, 2017, 30.