## Contributors: Ryan DeFever, Colin Targonski, Steven Hall ## Sarupria Research Group, Smith Research Group
## Clemson University
## 2019 Jun 22
--------------------------------------------------
## PLEASE CITE OUR WORK IF YOU USE THIS IN YOUR RESEARCH
Chem. Sci., 2019,10, 7503-7515
https://doi.org/10.1039/C9SC02097G
--------------------------------------------------
## REQUIREMENTS: - Python 3 [tested with 3.6.6] - Tensorflow [tested with 1.12.0] - Cython [tested with 0.29.3] - MDAnalysis (only required for reading .gro/.xtc files) [tested with 0.19.0] --------------------------------------------------
--------------------------------------------------
## OVERVIEW:
scripts/ --> Example scripts for training and evaluation models/ --> PointNet model in tensorflow utils/ --> Definition of datacontainer for PointNet datasets/ --> Example training data and example trajectory mda_custom/ --> Custom edits of MDAnalysis neighbor search --------------------------------------------------
--------------------------------------------------
## RELATED WORK:
This software would not be possible without the work of previous researchers. The neighbor search functionality is adapted from MDAnalysis (https://mdanalysis.org). The PointNet architecture is taken
from arXiv:1612.00593. Also see http://stanford.edu/~rqi/pointnet/
and https://github.com/charlesq34/pointnet. --------------------------------------------------
## COMPILING CUSTOM MDANALYSIS NEIGHBOR SEARCH:
cd mda_custom/nsgrid/ python setup.py build_ext --inplace --------------------------------------------------
--------------------------------------------------
## EXAMPLES: ### TRAINING POINTNET: python scripts/train.py --dataset datasets/lj-r2.0_scaled_shuffled_equal_samples.npy
--labels datasets/lj-r2.0_scaled_shuffled_equal_labels.npy
--weights PATH_TO_SAVE_WEIGHTS/weights Comments:
- See prepare_train.py for example of how samples and labels .npy files were prepared - WARNING, running train.py takes ~4 hours on one NVIDIA V100 GPU ### USING POINTNET TO CLASSIFY CRYSTAL STRUCTURES: python scripts/mda_cluster.py --weights PATH_TO_SAVE_WEIGHTS/weights
--nclass 4
--trjpath datasets/lj-seed
--cutoff 2.0
--maxneigh 43
--outname example Comments:
- nclass = 4, the network was trained to recognize four classes (liq,fcc,hcp,bcc) - maxneigh = 43, each point cloud has 43 points in training set. The number of points in each point cloud affects the network architecture.
- cutoff = 2.0, same as used for training
--------------------------------------------------
--------------------------------------------------
## GENERAL PROCEDURE FOR CRYSTAL STRUCTURE IDENTIFICATION: - Run simulations of pure phases at a range of (T,P) conditions - Pick a cutoff distance (targeting avg of ~30--50 points appears to give accurate classification) - Calculate mean and stdev number of points within cutoff distance for all phases - Choose the max number of points in each point cloud (mu+2*stdev seems to work well) - Extract training examples from pure phase simulations -- create .npy structures (see prepare_train.py for an example)
with training point clouds and labels in the SAME ORDER (a) Shape of numpy array for training point clouds should be [NEXAMPLES,NPOINTS,3] (b) Labels are one-hot encoded: Shape of numpy array for training labels should be [NEXAMPLES,NCLASSES] (c) Central atom for point clouds always translated to (0,0,0) (d) Points in point cloud scaled such that closest point to central atom is at a distance of 1.0 - Train PointNet (see train.py) - Use PointNet for classification (see mda_cluster.py) --------------------------------------------------