Official Repository for the paper Auto-resolving atomic structure at van der Waal interfaces using a generative model.
Note: if you want to install or run our codes, please cd to subfolders first.
[Paper]
Authors: Wenqiang Huang, Yuchen Jin, Zhemin Li, Lin Yao*, Yun Chen, Zheng Luo, Shen Zhou, Hongxia Wang, Jinguo Lin, Feng Liu, Zhifeng Gao, Jun Cheng, Linfeng Zhang, Fangping Ouyang*, Shanshan Wang* & Jin Zhang*
Unveiling atomic structures is significant for the relationship construction between microscopic configurations and macroscopic properties of materials. However, we still lack a rapid, accurate, and robust approach to automatically resolve complex patterns in atomic-resolution microscopy. Here, we present a Trident strategy-enhanced disentangled representation learning method (a generative model), which utilizes a few unlabeled experimental images with abundant low-cost simulated images to generate a large corpus of annotated simulation data that closely resembles experimental conditions, realizing simultaneous achievement of high quality and large volumes of the training dataset. A structural inference model is then trained via a residual neural network which can directly deduce the interlayer slip and rotation of diversified and complicated stacking patterns at van der Waals (vdWs) interfaces with picometer-scale accuracy across various materials (ReS2, ReSe2, and MoS2) with different layer numbers (bilayer and trilayers) and demonstrates robustness to defects, imaging quality, and surface contaminations. The framework can also identify pattern transition interfaces, quantify subtle motif variations, and discriminate moiré patterns that are undistinguishable in frequency domains. The high-throughput processing ability of our method helps discover a novel vdW epitaxy where various thermodynamically favorable slip stackings can coexist, demonstrating the machine learning contribution to the new knowledge emergence.
For optimal performance, we recommend a computer with the following specs:
-
RAM: 25+ GB
-
CPU: 4+ cores, 2.4+ GHz/core
-
GPU: 32 GB Tesla-V100 *1
The development version is tested on Linux operating systems. The developmental version has been tested on the following systems:
-
Linux: Ubuntu 20.04.5 LTS
-
Mac OSX:
-
Windows:
Mainly depends on the following python packages.
ase
argparse
collections
cv2
functools
subprocess
sklearn
numpy
PIL
torch
torchvision
transformers
If you are having an issue that you believe to be tied to software versioning issues, please drop us an issue.
- Clone this repo:
git clone https://github.com/dptech-corp/TED-Gen.git
cd code
Dataset of training Trident strategy-Enhanced-DRIT and Stacking-Pattern-Analyzer is available at zenodo
| Dataset | Size of compressed version | Download link |
|---|---|---|
| TED | ||
| ReS2 | 12.6 MB | download all |
| ReSe2 | 11.2 MB | download all |
| MoS2 | 28.9 MB | download all |
| SPA | ||
| ReS2 | 890.2 MB | download all |
| ReSe2 | 2.4 GB | download all |
| MoS2 | 27.2 GB | part01 part02 part03 part04 part05 part06 part07 part08 part09 |
You can also use the code we provide to generate simulated STEM images with or without noise.
Simulated STEM images were generated by the open-source 'incoSTEM' package in Computem. In Ubuntu 20.04, you may need to manually install libfftw3 - dev to use incoSTEM. Otherwise might not be able to generate simulation images.
apt-get update
apt-get -y install libfftw3-dev
Based on our script, you can generate simulated STEM images containing specified stacking patterns.You can generate an image with label by running the ipynb file in Computem_Gen subfolder.
You can also go to this page to obtain Computem software with GUI.
Check Computem_Gen for more detalis.
You can convert the noise-free stem images into realistic stem images using the Trident strategy-Enhanced-DRIT model we provide.
Before you start training or generating, please download the necessary model parameter files and datasets and configure them in the option.py file. If you want to train or test on your own dataset, change the --dataroot in the option.py file to the location of your own dataset.
For the format of the dataset, see example_data. The images in the label folder are not necessary for some downstream tasks and are only used in the generation step. For defect detection tasks, you may need to place your own label images in this folder.
Here to download the TED model parameters used in the paper.
| Model parameter files | Size | Download link |
|---|---|---|
| ReS2 | 780.3 MB | download |
| ReSe2 | 780.3 MB | download |
| MoS2 | 780.3 MB | download |
You can also run the following code to download all model parameters
cd Enhanced-DRIT
bash ./models/download_model.sh
You can train TED models using the datasets we provide or your own.
cd Enhanced-DRIT
python ./src/train.py
You can use the model parameters we provide or your own trained model parameters to obtain realistic STEM images. In the recommended configuration, the average time to generate an 1024*1024 image is about 1s.
cd Enhanced-DRIT
python ./src/test_transfer.py
In the following figure, the top panel is the simulation image input, and the bottom panel is the TED generated images.
Generation demo
Check Enhanced-DRIT for more detalis.
Once the dataset is ready, you can train a stacking pattern analysis model based on the code we provide. Alternatively, you can parse your material using existing model parameters.
You can train Stacking Pattern Analyzer models using the datasets we provide or your own.
cd Stacking-Pattern-Analyzer
python train.py
We recommend using our Online App to verify the parsing results.You can also test it in this ipynb file. In the recommended configuration, inference in ipynb takes an average of about 1s per image.
Online App
Several bilayer slip stacking ReS2 raw images and inference are showing in the following figure
Inference demo
Check this Stacking-Pattern-Analyzer for more detalis.
This project is licensed under the terms of the MIT license. See LICENSE for additional details.