asd-afm

Automated structure discovery in atomic force microscopy

This is a Pytorch version of Keras/Tensorflow model training code in https://github.com/SINGROUP/ASD-AFM. The original training procedure is not reproduced exactly, so some differences in the training outcomes may be expected.

Paper: B. Alldritt et. al, Automated structure discovery in atomic force microscopy, Sci. Adv., 2020.

Scripts:

• generate_data.py: Download molecule dataset and generate a dataset of AFM images and image descriptors.
• train.py: Train and test a model using the generated AFM image dataset.
• run_train.sh: Run the training script with preset parameters.
• predict_1S-camphor.py: Run prediction for experimental AFM images of 1S-camphor using a pretrained model.

Data:

• https://doi.org/10.5281/zenodo.10562769
  • afm_camphor.tar.gz: Experimental AFM images of 1S-camphor on Cu(111)
  • molecules.tar.gz: Two datasets of molecules (light and heavy) with point-charges.
• https://doi.org/10.5281/zenodo.10514470: Pretrained weights for the structure prediction model, one set of weights for each of the molecule datasets (light and heavy).