This is a proof of concept script of the paper https://www.nature.com/articles/s41597-020-0473-z The original paper is based on ASE_ANI, but here we use torchani instead For demo purpose here we only implemented the MD sampling described in the paper
pool: path to store initial dataset, and new data sampled during the AL loop
sampler.py: Wrapper functions to run MD sampling, and QM calculations after getting new structures Using ASE package to run MD with pre-trained ANI potential Using ORCA5 to run QM calculation
trainer.py: Function to train the model once new structures are generated during each AL epoch Different from the original ANI paper, we use LAMB optimizer here
utils.py: helper functions