Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Accelerating the Search for Superconductors Using Machine Learning

Suhas Adiga, Umesh V. Waghmare

TL;DR

This work addresses the challenge of predicting the superconducting critical temperature $T_c$ from chemical composition alone, circumventing the need for detailed crystal structures. It builds on Quantum Structure Diagram–inspired descriptors to engineer a composition-based feature set, and introduces a data-cleaning workflow (SuperCon-MTG) to produce a reliable training dataset. A two-step Random Forest framework—classification to identify superconductors and regression to estimate $T_c$—achieves strong performance, with SHAP analysis offering interpretable insights into feature importance. The approach successfully validates on recent literature, screens the Materials Project for potential superconductors, and demonstrates a path to accelerate discovery, while acknowledging limitations from non-structural data and data scarcity for non-superconductors.

Abstract

Prediction of critical temperature $(T_c)$ 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 $T_c$ of unconventional, higher $T_{c}$ 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 demonstrate that the principal component analysis of superconductivity data uncovers the clustering of various classes of superconductors. We use machine learning analysis and cleaned databases of superconductors to develop predictive models of $T_c$ of a superconductor using its chemical composition. Earlier studies relied on datasets with inconsistencies, leading to suboptimal predictions. To address this, we introduce a data-cleaning workflow to enhance the statistical quality of superconducting 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 $T_c$ as a function of descriptors motivated from Quantum Structure Diagrams. We demonstrate that this model generalizes effectively in reasonably accurate prediction of $T_{c}$ of compounds outside the database. We further employ our model to systematically screen materials across materials databases as well as various chemically plausible combinations of elements and predict $\mathrm{Tl}_{5}\mathrm{Ba}_{6}\mathrm{Ca}_{6}\mathrm{Cu}_{9}\mathrm{O}_{29}$ to exhibit superconductivity with a $T_{c}$ $\sim$ 105 K. Being based on the descriptors used in QSD's, our model bypasses structural information and predicts $T_{c}$ merely from the chemical composition.

Accelerating the Search for Superconductors Using Machine Learning

TL;DR

This work addresses the challenge of predicting the superconducting critical temperature from chemical composition alone, circumventing the need for detailed crystal structures. It builds on Quantum Structure Diagram–inspired descriptors to engineer a composition-based feature set, and introduces a data-cleaning workflow (SuperCon-MTG) to produce a reliable training dataset. A two-step Random Forest framework—classification to identify superconductors and regression to estimate —achieves strong performance, with SHAP analysis offering interpretable insights into feature importance. The approach successfully validates on recent literature, screens the Materials Project for potential superconductors, and demonstrates a path to accelerate discovery, while acknowledging limitations from non-structural data and data scarcity for non-superconductors.

Abstract

Prediction of critical temperature 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 of unconventional, higher 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 demonstrate that the principal component analysis of superconductivity data uncovers the clustering of various classes of superconductors. We use machine learning analysis and cleaned databases of superconductors to develop predictive models of of a superconductor using its chemical composition. Earlier studies relied on datasets with inconsistencies, leading to suboptimal predictions. To address this, we introduce a data-cleaning workflow to enhance the statistical quality of superconducting 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 as a function of descriptors motivated from Quantum Structure Diagrams. We demonstrate that this model generalizes effectively in reasonably accurate prediction of of compounds outside the database. We further employ our model to systematically screen materials across materials databases as well as various chemically plausible combinations of elements and predict to exhibit superconductivity with a 105 K. Being based on the descriptors used in QSD's, our model bypasses structural information and predicts merely from the chemical composition.
Paper Structure (20 sections, 11 equations, 11 figures, 7 tables)

This paper contains 20 sections, 11 equations, 11 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Descriptors from QSD's and Engineered features
  3. Methodology
  4. Preprocessing and Data Cleaning
  5. Machine Learning Workflow
  6. Results
  7. Unsupervised Learning: Principal Component Analysis
  8. Hyperparameter Tuning
  9. Classification
  10. Regression
  11. SHapley Additive exPlanations (SHAP) Analysis of the Random forest Model
  12. Random forest model performance with 5 Key descriptors
  13. Classwise performance of Machine Learning Model
  14. Validation of our model and Search for new Superconductors
  15. Superconductors from the recent literature
  16. ...and 5 more sections

Figures (11)

  • Figure 1: A schematics of data cleaning workflow employed in this study for preparation of SuperCon-MTG.
  • Figure 2: A schematic of Machine learning workflow
  • Figure 3: (a) Histogram of $T_c$ in SuperCon-MTG, where the first bin (at $T = 0$ K) has a width of 1 K, while all other bins have a width of 3 K with inset of distribution of $T_c$ for 5 largest classes of superconductors. (b) The distribution of compounds in the SuperCon-MTG dataset by class, including Cuprates, Alloys, Iron Chalcogenides, Iron Pnictides, Heavy-Fermion materials, Silicides, Borides, Oxides, Borocarbides, Tellurides, Germanides, Nitrides, Transition Metal Pnictides, Intercalated Graphite, Hydrides, Bismuthates, Elemental Superconductors (labeled as C2), A15 Compounds, Fullerides (labeled as C1), $\text{BiS}_2$-based materials, other Iron-based Compounds, Chevrel phases (labeled as C3), Dichalcogenides (labeled as C4), and Unclassified materials.
  • Figure 4: Clustering of superconducting materials in first two principal components (PC-1 and PC-2) space using the Principal Component Analysis (PCA). Distinct clusters corresponding to cuprates, elemental superconductors, and carbon-based compounds are evident and are shown as an inset in the plot.
  • Figure 5: Confusion matrices representing the classification performance for superconductors (labeled as SC) and non-superconductors (labeled as NSC) in the (a) Training and (b) Test datasets.
  • ...and 6 more figures