This app is based on training of more than 500 perovskites stuctures with more than 9000
conductivity data points at different atmospheres and temperatures.
About 7000 conductivity data was taken from the literature of:
Priya, P., Aluru, N.R. Accelerated design and discovery of perovskites with high conductivity
for energy applications through machine learning. npj Comput Mater 7, 90 (2021).
https://doi.org/10.1038/s41524-021-00551-3
The machine learning models for making this app are based on
XGBoost and neural network
with a customried loss function implimented using Pytorch. After training the two models
final prediction was taken
by the equily weighted model outputs. Featurelization of the compositions was based on pakages
of Pymatgen and
Matminer.
The raw data of parental oxides was taken from
Materials Project.
The acidity of parental oxides in smith scale are from the original paper and a prediction
from another machine learning model for oxides beyond the original paper. The parental oxides
were decomposed into 19 numerical representors (instead of more than 100 features in Priya's
paper). Eventhough, a small number of feature inputs were used in this app the accuray is
greatly enhanced from Priya's paper.
Since the input for this app is just the featurlised compositional data, it can be used for
predict conductivities with unknow stuctures. In the meantime, the the model may suffer form
lacking stuctural information of the materials. Due to the still limited data used for training
the prediction could be 1 order of magnitude different from real conductivity. In some case can
even be 3 mangnitudes difference.