Machine learning-driven design of cost-effective cutting tools: Composition optimization of transition-metal-doped WC

초록

In this study, we developed a machine learning (ML) framework based on 174 density functional theory (DFT)-calculated data points to predict the elastic properties of WC systems doped with 10 transition metals. A set of 13 descriptors, including valence electron concentration (VEC) and melting temperature (MT), were constructed and reduced to 8 non-redundant features. Among the models tested, Gradient Boosting achieved the best performance for bulk modulus (R-2 = 0.932), while Random Forest showed the highest accuracy for shear modulus (R-2 = 0.931). Shapley Additive exPlanations (SHAP) analysis identified VEC, MT, and electronegativity as key factors governing elastic behavior. Using the trained models, high-throughput screening of 1710 compositions revealed that Re doping dominates stiffness enhancement, with (W0.75Re0.25)C exhibiting the highest predicted performance (B = 397.5 GPa, G = 292.7 GPa). In addition, dual-doping strategies combining Re with other transition metals effectively reduced Re content while maintaining near-optimal properties. Furthermore, Re/Ru-free compositions such as (W0.917Cr0.056Mo0.028)C, (W0.889Cr0.111)C, and (W0.917Cr0.083)C achieved competitive performance (B > 380 GPa), suggesting viable cost-effective alternatives. These results demonstrate that ML-based screening can cost-effectively accelerate the design of high-performance WC-based carbides.

키워드