The following code accompanies the "Local structural features elucidate crystallization of complex structures" (https://arxiv.org/abs/2401.13765) preprint by M. M. Martirossyan, M. Spellings, H. Pan, and J. Dshemuchadse. This codebase utilizes the Steinhardt order parameter as well the descriptor and methods from Spellings and Glotzer (https://doi.org/10.1002/aic.16157). The methods utilized build local order metrics via unsupervised learning (dimensionality reduction and clustering) that allow for similar crystallographic sites in complex structures to be distinguished.
The following python libraries will need to be installed in order to run this code:
- numpy
- scipy
- pythia
- freud
- gsd
- gtar
- sklearn
- pickle
- loky
- matplotlib
This code can be used in conjunction with simulation data deposited to the Materials Data Facility (MDF) at the following site: https://doi.org/10.18126/wy01-4e11. Hyperparameters used for each structure are included as a training.json file.
The local_Ylms.py file contains all the functions for creating local descriptors and performing unsupervised learning to build local order parameters by clustering particles. The file also contains additional functions for plotting coordination number histograms as well as performing t-distributed stochastic neighbor embedding (tSNE), colored by data cluster.
The train_order_params.py file runs the functions in the local_Ylms.py file and applies them to the MDF data. This script will automatically generate folders grouped by descriptor type (Steinhardt parameter and Spellings' spherical harmonics descriptor) as well as by number of Gaussian mixture model (GMM) clusters.
For more information on the method for crystal identification (i.e., building global order parameters) by Spellings and Glotzer, look at: