Composition-Based Machine Learning for Screening Superconducting Ternary Hydrides from a Curated Dataset

TL;DR

The study addresses rapid, composition-based screening of high-$T_\mathrm{c}$ superconducting ternary hydrides under high pressure by training an ensemble of 30 XGBoost regressors on a curated dataset of ~2059 hydride entries. By leveraging 221 composition features and Bayesian feature filtering with nested cross-validation, the authors select a robust predictive model to screen over $18$ million A–B–H compositions across 100–300 GPa, using the lower bound of the ensemble's 95% CI for ranking. The results identify promising, hydrogen-rich candidates such as Ca–Ti–H, Li–K–H, and Na–Mg–H, and show that descriptors like ionization energy and atomic radius drive the learned composition-level trends in $T_\mathrm{c}$. While not incorporating explicit structural stability at this stage, the framework offers a practical, scalable tool for guiding experimental and computational efforts, with future work aimed at coupling with stability analyses and structure predictions to refine candidates.

Abstract

We present an ensemble machine-learning approach for composition-based, structure-agnostic screening of candidate superconductors among ternary hydrides under high pressure. Hydrogen-rich hydrides are known to exhibit high superconducting transition temperatures, and ternary or multinary hydrides can stabilize superconducting phases at reduced pressures through chemical compression. To systematically explore this vast compositional space, we construct an ensemble of 30 XGBoost regression models trained on a curated dataset of approximately 2000 binary and ternary hydride entries. The model ensemble is used to screen a broad set of A-B-H compositions at pressures of 100, 200, and 300 GPa, with screening outcomes evaluated statistically based on prediction consistency across ensemble members. This analysis highlights several high-scoring compositional systems, including Ca-Ti-H, Li-K-H, and Na-Mg-H, which were not explicitly included in the training dataset. In addition, feature-importance analysis indicates that elemental properties such as ionization energy and atomic radius contribute significantly to the learned composition-level trends in superconducting transition temperature. Overall, these results demonstrate the utility of ensemble-based machine learning as a primary screening tool for identifying promising regions of chemical space in superconducting hydrides.

Figures (4)

• Figure 1: Overview of the curated dataset of hydride superconductors used for model training. (a) Distribution of $T_\mathrm{c}$ values plotted against pressure and hydrogen per non-hydrogen atomic ratio. Higher $T_\mathrm{c}$ values are generally observed in hydrogen-rich compositions, particularly under high pressure. (b) Elemental occurrence heatmap showing the frequency of each element in the 2059 binary and ternary hydride compounds. Light elements, alkali metals, and alkaline earth metals are frequently represented, reflecting known experimental and computational focus areas.
• Figure 2: Evaluation of predictive fidelity and uncertainty estimation across 30 trained XGB models. (a) Parity plot comparing predicted and training $T_\mathrm{c}$ values, where error bars and color saturation indicate the standard deviation across ensemble predictions. (b) Correlation between prediction uncertainty (ensemble standard deviation) and absolute residuals, showing strong agreement with Pearson $r = 0.82$ and Spearman $\rho = 0.80$. The red dashed line shows the best-fit linear regression with a slope of 0.411, supporting the use of predicted variance as a confidence proxy.
• Figure 3: Feature importance analysis across 30 trained XGB models to assess interpretability and predictor diversity. (a) Cumulative importance curves based on gain, weight, and cover metrics aggregated over the ensemble. The gain curve shows a moderately distributed profile, with the top 16 features accounting for 50% of total importance, indicating no single dominant variable. (b) Top 10 features ranked by mean normalized gain. Error bars represent the standard deviation across models, and the label $N$/30 indicates how many models selected each feature. Commonly selected descriptors such as ionization energy and atomic radius suggest physically meaningful contributors to the predicted $T_\mathrm{c}$.
• Figure 4: Screening results for high-$T_\mathrm{c}$ ternary hydrides at 100, 200, and 300 GPa. Left panels: Ternary composition maps showing the lower bound of the 95% CI of predicted $T_\mathrm{c}$ values across the A--B--H space. Cyan stars indicate the highest-ranked compositions at each pressure, annotated with their formula and predicted lower CI value. Right panels: Heatmaps for each A--B elemental pair, with lower triangles showing the predicted 95% CI lower bound of $T_\mathrm{c}$ and upper triangles showing the standard deviation of ensemble predictions. These maps capture both performance and model uncertainty across the elemental search space. All top candidates are located in hydrogen-rich regions and exhibit high $T_\mathrm{c}$ with low uncertainty, highlighting their potential as robust superconductors.