The ASD-AFM repository contains the code for generating the datasets, training the machine learning models, and doing the configuration matching analysis for atomic force microscopy (AFM) images described in the article B. Alldritt et. al, Automated structure discovery in atomic force microscopy, Sci. Adv., 2020.
The dataset generation is done using the ProbeParticleModel AFM simulation code. The machine learning models are implemented in Tensorflow through the Keras API. The code is currently written in Python 2. At least the following Python packages are required:
- numpy
- matplotlib
- tensorflow-gpu
- keras
- pyopencl
- jupyter
Additionally, you need to have Cuda and cuDNN correctly configured on your system in order to train the models on an Nvidia GPU.
If you are using Anaconda, you can create the required Python environment with
conda env create -f environment.ymlThis will create a conda enviroment named tf-gpu with the all the required packages. It also has a suitable version of the Cuda toolkit and cuDNN already installed. Activate the environment with
conda activate tf-gpuTo create the datasets and train the models, run jupyter notebook in the repository folder, open the train_model.ipynb notebook, and follow the instructions therein.
Alternatively, run the script generate_data.py to generate the datasets and the script train_models.py to train the models.
The folder pretrained_weights holds the weights for pretrained models on the two datasets.
To make the configuration matching for experimental AFM images, run the experiment_1S-Camphor.ipynb notebook. You don't need to train a model to run this notebook.