Skip to content

This repository contains models to predict the compound disorder type

Notifications You must be signed in to change notification settings

lrcfmd/Disorder-type-prediction

Repository files navigation

Disorder-type-prediction

This repository contains the models to predict the disorder type of a compound (O-ordered, S-substitutional disorder, V-vacancies, M-mixed). This is classification done with the latest disorder script on ICSD (in which there is MC calculation of entropy and only six orbit disorder types: O, S, V, SV, VP, SVP. See corresponding paper draft for details)

Trained models are availible on the external hard drive. Models are trained on compounds with no hydrogen. However, there is data extracted for the compounds with hydrogens as well, so it is possible to eliminate this feature of the model by retraining the model on all compounds with "full structures" (which are defined as those having the same set of elements in structure and in composition).

Again there is a generic environment file, appologies

List of models:

RF + Magpie Balanced accuracy = 0.75, f1-score (macro) = 0.77, mcc = 0.73

CrabNet + Mat2Vec Balanced accuracy = 0.78, f1-score (macro) = 0.78, mcc = 0.75

Ensemble-10CrabNet + Mat2Vec Balanced accuracy = 0.80, f1-score (macro) = 0.81, mcc = 0.79

Confusion matrixes are in images folder

About

This repository contains models to predict the compound disorder type

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published