trellixvulnteam / graph-afm_zeeq

Machine learning molecule graphs from atomic force microscopy images.

Graph-AFM

Machine learning molecule graphs from atomic force microscopy images.

Paper: https://link.springer.com/article/10.1557/s43577-022-00324-3

Abstract: Despite the success of non-contact atomic force microscopy (AFM) in providing atomic-scale insight into the structure and properties of matter on surfaces, the wider applicability of the technique faces challenges in the difficulty of interpreting the measurement data. We tackle this problem by proposing a machine learning model for extracting molecule graphs of samples from AFM images. The predicted graphs contain not only atoms and their bond connections, but also their coordinates within the image and elemental identification. The model is shown to be effective on simulated AFM images, but we also highlight some issues with robustness that need to be addressed before generalization to real AFM images.

Setup and usage

This repository contains all the code and data used to achieve the results in the above paper. To run the code, first clone the repository:

git clone https://github.com/SINGROUP/Graph-AFM.git
cd Graph-AFM

Then, the Python environment required for running the code can be installed with Anaconda by using the provided environment.yml file:

conda env create -f environment.yml

and the installed environment can be activated with

conda activate graph-afm

The other environment file environment_exact.yml contains a dump of the exact list of packages used when training the model presented in the paper. Finally, run the script build.sh in the root of the repository to compile C extensions:

./build.sh

The model is trained on a dataset of simulated AFM images. The whole database is inconvenient to share directly due to its size, so we instead provide a database of molecule structures that can be used to generate the simulated AFM images. The AFM simulations are done using the ProbeParticleModel simulation code from Prokop Hapala. Navigate to the root of this repository and clone the ProbeParticleModel repository there with

git clone https://github.com/ProkopHapala/ProbeParticleModel.git
cd ProbeParticleModel
git checkout 363709ad3ed191422bfdcafea9ad6135e7906972

The database of AFM simulations can then be generated by using the script generate_data.py in the scripts directory:

cd scripts
python generate_data.py

The script will automatically download the molecule database and generate the simulations into a HDF5 archive file. The complete database takes ~150GB of disk space. The molecule database can also be downloaded directly from https://www.dropbox.com/s/z4113upq82puzht/Molecules_rebias_210611.tar.gz?dl=0.

The model can be trained using the script train.py, or train_distributed.py for multi-GPU training. Expect this to take a while. Training the model for 50 epochs took roughly 2 days on a system with 4 x Nvidia Tesla V100 32GB GPUs. After the model has been trained, statistics on test set predictions can be run using the script test.py.

Pre-trained weights for the model are provided in the directory pretrained_weights. See the scripts predict_examples.py and predict_random.py for examples on how to run predictions with the pre-trained model.