Scholarlink

<p>The rapid discovery of high-dielectric-constant materials is essential for the development of next-generation electronic, energy-storage, and communication technologies. However, conventional experimental characterization and first-principles calculations are computationally demanding for large-scale materials screening. In the present study, an explainable machine-learning framework was developed for dielectric constant prediction using 52,168 crystalline materials extracted from the Joint Automated Repository for Various Integrated Simulations (JARVIS-DFT) database. Structural, thermodynamic, electronic, and composition-derived descriptors were systematically investigated to evaluate their influence on dielectric behaviour. Pearson correlation analysis revealed positive correlations of dielectric constant with density (<i>r</i> = 0.396) and formation energy (<i>r</i> = 0.368), whereas band gap exhibited a negative correlation (<i>r</i> = − 0.379), consistent with established dielectric polarization theory. Three ensemble-based machine-learning algorithms, namely Random Forest (RF), Extra Trees (ET), and Extreme Gradient Boosting (XGBoost), were evaluated. The baseline XGBoost model achieved an R² value of 0.743, while the incorporation of composition-derived descriptors, including mean atomic mass and mean electronegativity, improved the prediction accuracy to R² = 0.789. To further evaluate model robustness, five-fold cross-validation and GridSearchCV-based hyperparameter optimization were performed. The cross-validation results demonstrated stable predictive performance (mean R² = 0.7533 ± 0.0031 for the optimized XGBoost model), confirming the generalization capability of the developed framework. SHapley Additive exPlanations (SHAP) analysis identified band gap, density, mean atomic mass, and mean electronegativity as the most influential descriptors governing dielectric behaviour, demonstrating the complementary roles of electronic structure and elemental chemistry. The novelty of the present study lies in the integration of structural, thermodynamic, electronic, and composition-aware descriptors within a unified explainable machine-learning framework, complemented by SHAP-based interpretation, systematic feature engineering, and rigorous model validation through five-fold cross-validation and GridSearchCV optimization. This integrated framework provides a transparent and computationally efficient approach for dielectric-property prediction and may serve as a useful foundation for future computational screening and experimental validation of high-k dielectric materials.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting dielectric constants of crystalline materials using explainable machine learning and composition-aware feature engineering

  • 期刊论文

  • 发表日期：

    2026年7月29日

查看全文
  • Devbrat Pundhir,
  • Ashok Kumar

摘要

The rapid discovery of high-dielectric-constant materials is essential for the development of next-generation electronic, energy-storage, and communication technologies. However, conventional experimental characterization and first-principles calculations are computationally demanding for large-scale materials screening. In the present study, an explainable machine-learning framework was developed for dielectric constant prediction using 52,168 crystalline materials extracted from the Joint Automated Repository for Various Integrated Simulations (JARVIS-DFT) database. Structural, thermodynamic, electronic, and composition-derived descriptors were systematically investigated to evaluate their influence on dielectric behaviour. Pearson correlation analysis revealed positive correlations of dielectric constant with density (r = 0.396) and formation energy (r = 0.368), whereas band gap exhibited a negative correlation (r = − 0.379), consistent with established dielectric polarization theory. Three ensemble-based machine-learning algorithms, namely Random Forest (RF), Extra Trees (ET), and Extreme Gradient Boosting (XGBoost), were evaluated. The baseline XGBoost model achieved an R² value of 0.743, while the incorporation of composition-derived descriptors, including mean atomic mass and mean electronegativity, improved the prediction accuracy to R² = 0.789. To further evaluate model robustness, five-fold cross-validation and GridSearchCV-based hyperparameter optimization were performed. The cross-validation results demonstrated stable predictive performance (mean R² = 0.7533 ± 0.0031 for the optimized XGBoost model), confirming the generalization capability of the developed framework. SHapley Additive exPlanations (SHAP) analysis identified band gap, density, mean atomic mass, and mean electronegativity as the most influential descriptors governing dielectric behaviour, demonstrating the complementary roles of electronic structure and elemental chemistry. The novelty of the present study lies in the integration of structural, thermodynamic, electronic, and composition-aware descriptors within a unified explainable machine-learning framework, complemented by SHAP-based interpretation, systematic feature engineering, and rigorous model validation through five-fold cross-validation and GridSearchCV optimization. This integrated framework provides a transparent and computationally efficient approach for dielectric-property prediction and may serve as a useful foundation for future computational screening and experimental validation of high-k dielectric materials.