Predictive Modeling of Critical Temperatures in Superconducting Materials

Area of Science:

• Materials Science
• Computational Chemistry
• Data Science

Background:

• Quantitative structure-property relationships (QSPR) are crucial for understanding material properties.
• Previous superconductor critical temperature models were based on datasets with significant duplicate entries (27%).
• Accurate prediction of critical temperature is essential for developing new superconducting materials.

Purpose of the Study:

• To develop stable and accurate predictive models for the critical temperatures of inorganic superconductors.
• To address data quality issues in existing superconductor datasets.
• To identify key physicochemical attributes influencing critical temperature.

Main Methods:

• Data cleaning and preprocessing to remove duplicate entries.
• Application of machine learning techniques: multiple linear regression (MLR), gradient boosting decision trees (XGBoost), neural networks (NN), and random forests (RF).
• 10-fold cross-validation to evaluate model performance using R-squared (R2) and root-mean-square error (RMSE).

Main Results:

• Developed an XGBoost model with R2 = 0.924 and RMSE = 9.336, comparable to previous models but using a cleaned dataset.
• Identified thermal conductivity as the most influential variable for predicting critical temperature.
• Achieved comparable predictive accuracy with simpler, more interpretable parameters.

Conclusions:

• Data quality significantly impacts the reliability of predictive models for superconductor critical temperatures.
• XGBoost and other machine learning models can provide accurate and generalizable QSPR models for superconductors.
• Physicochemical attributes, especially thermal conductivity, are vital predictors of superconducting behavior.