Clustering Superconductors Using Unsupervised Machine Learning

TL;DR

This paper explores the use of various unsupervised machine learning algorithms to identify clustering structures in superconducting materials datasets, combining computational methods with domain knowledge for improved results.

Contribution

It demonstrates the effectiveness of combining machine learning techniques with human expertise to cluster superconductors and highlights the importance of staged clustering for detailed substructure detection.

Findings

Machine learning can match or surpass human clustering performance.

t-SNE is the most effective for visualization.

Staged clustering reveals finer subcluster structures.

Abstract

In this work we used unsupervised machine learning methods in order to find possible clustering structures in superconducting materials data sets. We used the SuperCon database, as well as our own data sets complied from literature, in order to explore how machine learning algorithms groups superconductors. Both conventional clustering methods like k-means, hierarchical or Gaussian mixtures, as well as clustering methods based on artificial neural networks like self-organizing maps, were used. For dimensionality reduction and visualization t-SNE was found to be the best choice. Our results indicate that machine learning techniques can achieve, and in some cases exceed, human level performance. Calculations suggest that the clustering of superconducting materials works best when machine learning techniques are used in concert with human knowledge of superconductors. We also show that in order to resolve fine subcluster structure in the data, clustering of superconducting materials should be done in stages.