This github is for the project: Intermetallic Crystal Nanoparticle Model for Predicting Surface Energy and Synthesizability for Catalyst Discovery.
We developped intermetallic surface energy database, featuring 12,553 entries and 344,200 DFT single points. Based on the database, we trained a machine learning force field model to achieve 10^5 times speedup in DFT relaxations for surface energy calculations. The model aims to provide fast and accurate surface exposure and synthesizability information of intermetallic nanoparticles for large scale catalyst screening and catalyst discovery
The code is developed and tested on Windows 10/11 OS. This section provides the instructions to install the required packages/dependencies and download the dataset.
Clone the repository and install the required packages using the following commands: The installation would take less than 5 minutes.
git clone https://github.com/Long1Corn/SurFF.git
cd SurFF
bash install.shRequired - Download files of the dataset, the trained models, and results:
python download_data.py- Currently, please manually download the dataset from the link provided in the download_data.py file.
- To simply run the prediction example, you can just download
SurFF_CoreDataFiles.zip(130MB). - To reproduce the complete experiment, you should download the complete dataset
SurFF_DataFiles.zip(4.6GB).
- The relaxation trajectory is provided in Data/Surface/Traj folder.
- We also provide the calculated surface energy in Data/Surface/Surface_Energy folder.
- The coresponding structure files are provided in Data/Surface/Structure folder.
Summary of intermetalic surface datasets generated in this work.
| Dataset | Surfaces | DFT single point | DFT time (cpu-hr) | Notes | File Location |
|---|---|---|---|---|---|
| Train set | 9,676 | 262,984 | 155,612 | Mode training; | train_lmdb |
| AL test set | 1,865 | 49,295 | 28,371 | Evaluate accuracy while dataset generation; Test surface energy accuracy; | test_al_lmdb |
| ID test set | 327 | 7,995 | 2,642 | Test surface energy accuracy; Test surface area accuracy; | test_id_lmdb |
| OOD test set | 685 | 23,926 | 7,576 | Test surface energy accuracy; Test surface area accuracy; | test_ood_lmdb |
| Pred set | 144,191 | - | - | Predict surface energy and area for intermetallic crystals; | pred_lmdb |
- All the trained model are provided via saved_model folder.
Relaxation Results
| Model | Dataset | MAE (meV/Ų) | Fraction | Energy | L | M | H | Top3 | Top5 | GPU Time (hr) |
|---|---|---|---|---|---|---|---|---|---|---|
| Base | OOD | 10.5 | 0.074 | 0.669 | 0.633 | 0.661 | 0.744 | 0.666 | 0.810 | 0.27 |
| Base | ID | 2.9 | 0.032 | 0.734 | 0.772 | 0.627 | 0.803 | 0.800 | 0.810 | 0.12 |
| Base | AL | 3.8 | - | - | - | - | - | - | - | 0.69 |
| Finetune | OOD | 6.8 | 0.080 | 0.675 | 0.666 | 0.605 | 0.744 | 0.700 | 0.820 | 0.26 |
| Base | Pred | - | - | - | - | - | - | - | - | 115 |
Single Point Results
| Model | Dataset | Force MAE | Force Cos | Energy MAE |
|---|---|---|---|---|
| Base | OOD | 0.1949 | 0.3039 | 2.3473 |
| Base | ID | 0.0334 | 0.5594 | 0.3229 |
| Base | AL | 0.0472 | 0.5338 | 0.4199 |
| Finetune | OOD | 0.0902 | 0.3814 | 1.2926 |
- By applying SurFF, we provide the surface energy and exposure information for over 6,000 intermetallic crystals and 14 thousands surfaces. The results are provided in results/all folder.
- Example predictions and comparison with experimental observations.
Example of using the model to predict surface exposure and area for intermetallic crystals. The sample codes would take less than 15 minutes to run, exclude the time for downloading the dataset.
We also provide a web UI for users to predict surface energy and area for intermetallic crystals.
To use the web UI, please install the required packages and download the core data files. Then run the following command:
cd ocp
python app.py Then simply click the browser link provided in the terminal. A guide for using the web UI is provided in Web UI Guide.
To reproduce the complete experiment, please download the complete dataset SurFF_DataFiles.zip(4.6GB).
The model training configuration and hyperparameters are provided in config folder. Please refer to the SI of the paper for more details.
Distributed under the MIT License. See LICENSE.txt for more information.
Yin Jun - yinjun98@u.nus.edu
Chen Honghao - chh22@mails.tsinghua.edu.cn
