Title: Prediction and Virtual Screening of Elastic Properties in High-Entropy Carbides Using Composition-Based Descriptors and Machine Learning

Journal: Ceramics International

Year: 2026

Impact factor: 6.0

Abstract:

High-entropy carbides (HECs) have attracted considerable attention for extreme-environment applications owing to their compositional flexibility and high-temperature mechanical properties. Among their key properties, Young’s modulus is closely related to stiffness and thermo-mechanical reliability, yet the vast compositional space of HECs makes exhaustive experimental characterization and first-principles screening impractical. Here, we present a composition-based machine-learning framework for predicting Young’s modulus and identifying its key descriptors. Using a high-fidelity density functional theory (DFT)–derived dataset of 146 HEC compositions, random forest (RF), support vector regression (SVR), and artificial neural network (ANN) models were trained. The ANN showed the best test-set performance (R2 = 0.88), while the SVR achieved the best 5-fold cross-validation performance (R2 = 0.87). Feature-importance and SHapley Additive exPlanations (SHAP) analyses revealed that valence electron concentration (VEC)-related descriptors are the most influential factors, with melting temperature- and atomic radius-related descriptors providing complementary contributions. Virtual screening of 6426 quinary compositions identified high-modulus candidates, with (Nb0.25Ta0.25W0.20Ti0.15Hf0.15)C ranked highest (487.44 GPa). These results demonstrate an interpretable and efficient route for data-driven discovery of high-stiffness HECs.