This repository contains a collection of notebooks and data that investigates correlations between vibrational properties and diffusive properties in solids. More background can be obtained from the chapter Do phonons act as an effective descriptor for fast-ion transport? of my thesis.

The output data obtained by myself running the notebooks presented here can be found in the data repository

The phonon data is obtained from the phonon calculation database created by Atsushi Togo. Please refer to (https://github.com/WMD-group/phononDB) to query the phonon data.

The notebook 01_get_data.ipynb contains functions required to organise the phonon density of states (DOS) data from PhononDB into a single folder that can be easily worked with later on, as well as functions to calculate phonon descriptors obtained from the DOS data: phonon band-centre, relative spread, DOS first peak, and DOS spectrum featurisation.

After running 01_get_data.ipynb, a dataframe data.csv, and feature vectors of the DOS (viball_feature.npy, vibli_feature.npy, vibtot_feature.npy) are obtained. data.csv includes composition data as well as descriptor data for each Li-containing material found in the PhononDB database.

Then, in the 02_labelling_conductivies notebook, conductivity labels are added to the dataframe using digitized_data_for_SSEs.csv, a database obtained from (https://github.com/FALL-ML/materials-discovery), giving the updated dataframe labelled_data.csv.

The labelled data can be visualised in 3aa_visualisation.ipynb, where fast-ion conducting outliers and candidates can be identified (first_peak_candidates.csv, pbc_tot_rt_candidates.csv, pbc_diff_unknown_*.csv). Unsupervised clustering is also attempted in 3ab_clustering.ipynb and new candidates are identified (circle_dos_fv.csv).

M3GNet simulations are then run for all the candidates and more. The data is stored in stoich.csv. The results are investigated in 3ac_candidates_investigation.ipynb.

M3GNet simulations were finally run for all the materials in the PhononDB database by Kasper Tolborg (data_kasper_full.csv). Direct correlations between phonon descriptors and diffusivity are investigated in 3b_firect_correlations.ipynb.

An attempt to reduce the dimensionality of the DOS feature vectors using a (Variational) AutoEncoder can be found in aa_AE.ipynb and bb_VAE.ipynb.

The notebook b_full_dataset_labelling.ipynb can be used to create a dataframe, data_labelled_kasper.csv, that includes all the dataset with phonon descriptors and diffusivity data.

A composition-structure-phonon cross-correlation was also attempted. Composition and structure feature vectors can be obtained using ca_composition_featuriser.ipynb and cb_structure_featuriser.ipynb, respectively. To run the cd_csv_corr.ipynb notebook, the DOS feature vectors must be first converted to .csv files using the notebook cc_npy_to_csv.ipynb.

This work builds upon the work of Amelia Hu realised as part of a UROP summer internship. Her original work can be found at (https://github.com/AmeliaHu0920/urop-project). The internship was co-supervised by Anthony Onwuli and myself.

The data data_kasper_full.csv and stoich_data.csv was obtained by Kasper Tolborg using M3GNet.