Towards In Silico Mining for Superconductors -- Cutting the Gordian Knot

TL;DR

This paper presents a machine learning approach using random forest regression to predict the critical temperature of superconductors from electronic band structures, bridging experiment and theory without assuming a specific microscopic mechanism.

Contribution

It introduces a novel method combining DFT band structures with machine learning to predict superconductor properties across material classes.

Findings

Validated on A15 intermetallics including V, Nb compounds.

Features away from the Fermi level are crucial for predicting superconductivity.

The method can classify and predict new superconducting materials.

Abstract

A random forest regression based supervised machine learning method to predict experimental critical temperature of superconductivity from the electronic band structure, as obtained from Density Functional Theory, is demonstrated. This complementarity between experiment and theory draws inspiration from the merging of Kohn-Sham and Bogoliubov-De Gennes equations [W. Kohn, W, EKU Gross, and LN Oliveira, Int. J. of Quant. Chem., 36(23), 611-615 (1989)]. Features in the Kohn-Sham Density Functional Theory band structure away from EF becoming decisive for the superconducting gap demonstrates this divide-and-conquer physical understanding. Not committing to any microscopic mechanism for the SC at this stage, it implies that in different classes of materials, different electronic features are responsible for the superconductivity. However, training on known members of a class, the performance of new members may be predicted. The method is validated for the A15 materials, including both binary A3X and ternary A6XY intermetallics, A=V, Nb, demonstrating that the two do indeed belong to the same class of superconductors.