A framework to learn hierarchical motifs in disorder-order transitions: mapping structural motifs in supercooled water and ice polymorphs

This repository contains the dataset and code for the preprint: Mapping Hierarchical Motifs in disorder-order transitions: learning the hierarchy of structural motifs in supercooled water and ice polymorphs In the order mentioned in the main text, the subfolders include: 1. 'Nucleation-Trajectory' folder contains the code for training the VAE model 'model-training': The nucleation trajectory dataset comprising 141 frames ('nucleation-trajectory-frames.tar.gz'), as subsampled from a 1.2 micro-second homogenous nucleation trajectory The 'nucleation-trajectory-141frames-metadata.csv' maps the time steps are each snapshot There is also the sample scripts used to obtain the atomic descriptors: 'featurize-steinhardt.py': Steinhardt descriptors (using pyscal3 package), 'featurize-soap.py': SOAP (dscribe) 2. 'Variational-Autoencoder' folder contains the code for training the VAE model 'model-training': The VAE training script "train_vae.py" and two utility scripts - util.py (for data preprocessing) and 'model_concatdecoder.py' (for defining model decoder) The model archectiture is shown in 'model-archectiture.txt' The 'final-model' contains the final VAE model 'VAE.final.pth' and script for evaluating the VAE 'evaluation.py' We note that the full trajectory and descriptors are not publicly hosted due to file-size constraints; they are available upon reasonable request 3. 'Ice-Polymorph' folder contains the data for analysis of hypothetical ice polymorphs The codes for computing the structural descriptors of ice polymorphs are provided in 'compute-structural-descriptors' The structure and metadata files of nearly 252,000 candidate ice polymorphs is in 'candidateIcePolymorphs-252000.tar.gz' The heterogenous nucleation trajectories of four candidate ice polymorphs are provided in 'heterogenous-nucleation-four-candidates' as indicated by their label and ID.