PerovskiteSynthesizability_Manuscript2021

In hope to quantify the synthesizablility of virtual crystals, we implemented transfer learning and positive-unlabeled (PU) learning to predict synthesizability. This repository contains code to reproduce the result in our manuscript.

Developers

Geun Ho Gu (ggu@udel.edu) <- Current maintainer

Dependencies

• Python3
• Scikit-learn
• Numpy
• Pytorch
• Pymatgen
• Pytorch_scatter (https://github.com/rusty1s/pytorch_scatter)

Installation

1. Clone this repository:

git clone https://github.com/kaist-amsg/PerovskiteSynthesizability_Manuscript2021

Data Availablility

In the ./data folder, we have all raw crystal data in cif format. The source of the cif data (e.g. MP, OQMD, AFLOW ids) are in the cif_sources.json. The provided python codes contain preprocessing, model training, and synthesizability prediction.

Guide

1PU_Split.py splits the data (splits_Perov_All.json) in accordance to the inductive PU learning method. 2Train.py trains 100 PU learning model. The model weights trained with MP data is in base_weights, which are then transferred and retrained, weight of which are saved in weights folder. 3Predict.py makes prediction to all the data, which is summarized in the file "prediction.csv." For making new prediction, use 4New_Predict.py, and put test data in data4prediction folder. The result will be summarized in prediction.csv.

• ./data folder contains crystal data (*.cif), elemental combination of each crystal file (Perov_All_ABC.json), and the source of each cif file (cif_sources.json), and the icsd label (id_prop.csv).
• ./gcnn folder contains the python module for data preprocessing and model construction.
• ./predict folder contains model prediction in json file for all the crystal data.
• ./tests folder contains prediction result for the held-out test sets.
• ./base_weights contains model weights trained with MP data, and "./weights" contains retrained paramaters after trasnferring. The CL score prediction result is in the SI of the manuscript, but it is also tabulated in the "prediction.csv."

Publications

If you use this code, please cite:

Geun Ho Gu, Jidon Jang, Juhwan Noh, Aron Walsh & Yousung Jung, Perovskite synthesizability using graph neural networks, npj Computational Materials volume 8, Article number: 71 (2022)