This repository contains models predicting whether each element in the composition formula is ordered/both ordered and disordered/disordered (3 classes)
The latest model is a modified CrabNet. Also it would be reasonable to use Roost for this task, but I think I have lost the old code for that, and did not write the new one recentely..
CrabNet + Mat2Vec
Per element balanced accuracy = 0.79, per-element f1-score = 0.79, per-element mcc = 0.77
Compound accuracy (compound is classified correctly if all elements in it are classified correctly) = 0.71
Compound accuracy for ordered compounds = 0.88
Compound accuracy for disordered compounds = 0.59
Ensembling does not help
I think that we also can train model which has only two classes -- ordered and disordered (meaning that it has at least one disordered orbit) -- it should have higher scores