Assessing data-driven predictions of band gap and conductivity for transparent conducting materials

This repository contains python code accompanying the paper Assessing data-driven predictions of band gap and electrical conductivity for transparent conducting materials.

Installation

1. Clone this repository:

   git clone https://github.com/fedeotto/tcms
   

2. Install a new conda environment from env.yml:

   conda env create -f env.yml
   

3. Activate the new environment:

   conda activate tcms
   

Change the .env.template file to .env , modifying the necessary paths accordingly.

Usage

Data

Due to licensing restrictions associated with MPDS data and confidentiality agreements tied to funding, we are unable to publicly release the full experimental datasets used in this study (band gap and conductivity). However, we provide representative data that can be used as demonstration to run the proposed pipeline.

First, you can download the updated version of UCSB dataset at the following link (te_expt.xlsx) and move it into the data folder. Then, run the script prepare_data.py via:

python -m data.prepare_data

This will automatically create conductivity.xlsx and bandgap.xlsx datasets in the data folder that can be used to run the code. Band gap data is automatically extracted using matminer (matbench_expt_gap).

Trained models

We provide access to trained models (CrabNet and Random Forest) on the full data presented in the paper at the following link (GDrive).

Screening new materials from custom material lists

Trained models can be used to predict electrical conductivity and band gap from arbitrary chemical compositions. You can reproduce the results illustrated in Table 4 of the paper using trained models via

python main.py action=screen model=crabnet ++screen.screen_path=datasets/tcms.xlsx

Jupyter Notebooks

We include jupyter notebooks illustrating the analysis presented in the paper:

1. att_coeff.ipynb: contains the analysis relative to the attention coefficients of CrabNet in the task of leave-one-TCM-out described in the paper.

2. bandgap_prediction.ipynb: illustrates the comparison on the task of band gap prediction across different settings, in particular Random Forest, CrabNet with no fine-tuning and CrabNet fine-tuned on Materials project band gap data.

3. parity.ipynb contains additional visualization and plotting corresponding to electrical conductivity and band gap predictions.

Other usage

It is also possible to evaluate/fit new machine learning models on available data. For example, to fit a new CrabNet model on conductivity data you could simply do it via:

python main.py action=fit model=crabnet data=conductivity