Using unsupervised machine learning to predict critical temperatures of superconductors

We use the superconductors from the SuperCon database to
construct element vectors and then perform unsupervised learning
of critical temperatures. Only the chemical composition of
superconductors is used in this procedure. No physical predictors
(neither experimental nor numerical) of any kind are used. We
archive R²=0.93 which is comparable and in some cases higher
then similar estimates using other artificial intelligence
techniques. Based on this machine learning model, we predict
several new superconductors with high critical temperatures. We
also discuss the factors that impede the learning process and
suggest possible ways to fix them.