Accelerating the Search for Superconductors Using Machine Learning

Suhas Adiga, Umesh V. Waghmare · 2025

Read Accelerating the Search for Superconductors Using Machine Learning by Suhas Adiga, Umesh V. Waghmare (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-11964.txt.

Introduction

Suhas Adiga and Umesh Waghmare speed the hunt for superconductors with machine learning, in this 2025 paper. Guessing the temperature where a material turns superconducting is a standing puzzle. Their model reads only the chemical recipe, skipping structure entirely.

First the team cleaned the known superconductor records, dropping repeats and fixing clashes. Pattern analysis showed material classes cluster in the cleaned data. A random forest then learned to predict the key temperature from recipe-based traits. The model held up on compounds outside its training set. A wide screen flagged one new compound as a high-temperature candidate near 105 K.

Catalog record

CREATORSuhas Adiga, Umesh V. Waghmare
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE54 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-11964

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Accelerating the Search for Superconductors Using Machine Learning Suhas Adiga, Umesh V. Waghmare Abstract Prediction of critical temperature ( T c T_{c} ) of a superconductor remains a significant challenge in condensed matter physics. While the BCS theory explains superconductivity in conventional superconductors, there is no framework to predict critical temperature of unconventional, higher T c T_{c} superconductors. Quantum Structure Diagrams (QSD) were successful in establishing structure-property relationship for superconductors, quasicrystals, and ferroelectric materials starting from chemical composition. Building on the QSD ideas, we first demonstrate that the principal component analysis of superconductivity data naturally uncovers the clustering of various classes of superconductors. Earlier machine learning studies generally relied on datasets with inconsistencies and incomplete information, leading to a suboptimal predictions of superconductivity. To address this, we introduce a data-cleaning workflow to enhance the statistical quality of superconducting material databases by eliminating redundancies and resolving inconsistencies. With this improvised database, we apply a supervised machine learning framework and develop a Random forest model to predict superconductivity and critical temperature as a function of descriptors motivated from QSD. Our ML model demonstrates robust performance, achieving an R 2 R^{2} score of 0.87 and a mean absolute error of 5.6 K. We demonstrate that this model generalizes effectively in reasonably accurate prediction of T c T_{c} of compounds outside the database. We further employ our model to systematically screen materials across multiple materials databases as well as various chemically plausible combinations of elements and predict Tl 5 ​ Ba 6 ​ Ca 6 ​ Cu 9 ​ O 29 \mathrm{Tl}_{5}\mathrm{Ba}_{6}\mathrm{Ca}_{6}\mathrm{Cu}_{9}\mathrm{O}_{29} to be a superconductor with a T c T_{c} ∼ \sim 105 K. Being based on the descriptors used in QSD’s, our model bypasses structural information and predicts T c T_{c} merely from the chemical composition, and hence is quite efficient and powerful in the search for new superconductors. Keywords Superconductivity, Machine learning, Critical temperature I Introduction The discovery of superconductivity in mercury (Hg) was serendipitous, occuring below a critical temperature of 4.2 K and first observed by Onnes in 1911 [ 59 ] . Superconductors are materials that conduct electric current with zero resistance below the critical temperature ( T_{\textnormal{c}} ). They are used in number of applications in Magnetic Resonance Imaging (MRI), Maglev trains, particle accelerators, energy transmission and as qubits for next-generation quantum computers. Shortly after the discovery of superconductivity in elements like Hg and Pb, it was also observed in alloys [ 75 ] , A15 compounds [ 31 ] , cuprates [ 9 ] , iron-based pnictides [ 39 ] , and metal hydrides [ 69 ] . Currently, the highest T_{c} achieved at ambient pressure is 138 K in Hg-based cuprates [ 70 ] , which increases to 160 K under a pressure of 23 GPa [ 24 ] . Ashcroft argued that metallic hydrogen, under high pressure, could exhibit even higher critical temperature [ 5 ] . To date, the highest T_{\textnormal{c}} is observed in La \textnormal{H}_{10} , which shows near-room-temperature superconductivity at 260 K, albeit under an extremely high pressure of 190 GPa [ 76 ] . Superconductors are broadly of two types: Conventional and Unconventional superconductors. Conventional superconductors are those where the mechanism of superconductivity is based on electron-phonon coupling within the BCS Theory [ 8 ] , where phonons provide attractive interactions between two electrons that form a Cooper pair. Compounds like intercalated graphite (e.g. \text{C}_{6}\text{Yb} [ 86 ] ), A15 compounds (e.g. \text{V}_{3}\text{Si} [ 31 ] ), elemental compounds (e.g. Hg [ 59 ] ), and metal hydrides (e.g \text{CeH}_{9} [ 47 ] ) belong to this class. Eliashberg Theory, an extension of the BCS theory that incorporates the full electron-phonon interaction [ 22 ] , complemented by the empirical equation proposed by McMillan [ 55 ] and Allen-Dynes [ 3 ] estimates the critical temperature of superconductors based on electron-phonon coupling and average phonon frequency. This formalism can be readily used within the framework of first-principles Density Functional Theory to predict critical temperature [ 19 ] of conventional superconductors. Among the known conventional superconductors at ambient pressure, the highest critical temperature is observed in Magnesium Diboride ( \textnormal{MgB}_{2} ) with a critical temperature of 39 K [ 57 ] . High temperature superconductors like cuprates, iron-based compounds are unconventional superconductors, where there are various mechanisms like hole theory [ 33 ] , resonating valence bond theory [ 4 ] . However, these mechanisms do not quantitatively predict critical temperature. While achieving superconductivity at room temperature is a dream of condensed matter physicists, developing a predictive model of critical temperature remains a serious challenge. The high computational costs for screening thousands of compounds using Density Functional Theory (DFT) makes it resource intensive. Additionally, DFT based methods are limited to predict critical temperature in a weakly correlated conventional systems. This hinders the exploration for materials with potential for room temperature superconductivity. Statistical methods have been powerful tools, even before the conceptualization of machine learning. One such statistical approach was developed in the 1980’s by J.C. Phillips. It involved using Quantum Structural Diagrams [ 65 ] to establish structure-property relationships based on chemical composition. Many attempts have been made since 2018 to estimate critical temperature using machine learning. Stanev et al. [ 77 ] used 145 Magpie descriptors and 26 Aflow features based on elemental composition to predict critical temperature using Random forest method. Hamidieh’s [ 30 ] work presented a data-driven statistical model to predict critical temperature of superconductors using only the chemical formulas of the materials. This model extracted 81 features based on the elemental properties. Gashmard et al. [ 26 ] used 322 atomic features generated from Soraya package in Python to predict critical temperature of superconductors. Zhang et al. [ 89 ] used machine learning with structural descriptors and electronic properties from the Aflow database to predict the critical temperature of superconductors. Jung et al. [ 37 ] developed a framework that uses gradient-boosting and statistical feature selection to predict critical temperature. Trezza et al. [ 78 ] developed a machine learning classification framework to distinguish superconductors above and below temperature thresholds of 15 K (for benchmarking) and 35 K (for mixed-feature optimization). Their model was based on the Matminer descriptors [ 84 ] and incorporated several techniques, including invariant group identification and SHAP for feature reduction. Classification was performed using entropy-based Quasi Equilibrium Grids and Extra Trees. In their later work [ 79 ] , they extended this approach to address sampling bias between specialized databases like SuperCon [ 58 ] and general-purpose databases like the Materials Project (excluding SuperCon entries) [ 35 ] , thereby enabling validation of T_{c} predictions, even though the latter may include materials not tested for superconductivity. Konno et al. [ 42 ] introduced a novel deep learning model that learns to interpret the periodic table, rather than relying on manually engineered features. This model was trained on the materials from SuperCon and Crystallographic Open Database (COD). Kaplan et al. [ 40 ] employed a convolutional neural network (CNN) to analyze the chemical compositions of superconductors without augmenting non-superconductors in their model. Pereti et al. [ 64 ] utilized a Deep Set neural network that processes chemical compositions as sets, thus eliminating the

Read the complete text