sunatthegilddotcom/cubic-perovskite-workfunction

Cubic-Perovskite-WorkFunction

This repository contains the dataset perovskites' work functions using first principles and the electronic descriptors to perform data-driven analysis. This work aims to identify the electronic factors and how they control perovskite oxides' work functions.

The following paper describes the details of the analysis:

Data-driven analysis of the electronic factors controlling the work functions of perovskite oxides

Table of Contents

• Prerequisites
• Usage

Getting Started

Prerequisites

• scikit_learn
• Numpy
• pandas
• joblib
• biokit
• pdpbox
• matplotlib
• seaborn

biokit is used for plotting correlation matrix, pdpbox is used to perform partial dependence analysis (PDP). The prerequisites can also be installed using pip.

pip install -r requirements.txt

Usage

The analysis of the perovskites' work functions contain three main steps, the help messages can be find using:

python correlated_matrix.py -h
python rfecv.py -h
python model_analysis.py -h

The dataset contains the original 38 electronic features and the first principles-derived work functions is in the data folder, the definition of the features can be found in our manuscript: Data-driven analysis of the electronic factors controlling the work functions of perovskite oxides

1. remove correlated features, default by enumeration method and dumping the plot:

   python correlated_matrix.py -e -p

   This step creates new dataset perovskite_wf_data_uncorr.csv in data directory.

2. recursive feature elimination with cross-validations, hyperparameters tuned at each step, for example:

   python rfecv.py --term A --rfecv

   This perform the recursive feature elimination for A-termination of perovskites based on random forest regressor, the hyperparameter is selected at each step. The optimized features and transformed dataset are saved in rfecv_results and data directory, respectively.

3. Analysis of the model's performance, with partial dependence analysis.

   python model_analysis.py --term A --pdp_1d --pdp_2d

   --pdp_1d and --pdp_2d flag will perform the partial dependence analysis, available feature names will be displayed.

rfecv and pdp results can be plotted using plotter.py in plot_tools directory:

python plotter.py --rfecv --pdp

the obtained figures are stored in rfecv_results and results directory, respectively.

The trained models for AO and BO₂-terminated interfaces are stored in model_checkpoints.