The microstructure of molten marks changes according to ambient temperatures, when a short circuit occurs. Investigation of microstructural changes is important for understanding the properties of copper and examining the cause of a fire. In this study, the boundary characteristics and grain-size distribution of molten marks—primary-arc beads (PABs), which short-circuited at room temperature (25 °C), and secondary-arc beads (SABs), which short-circuited at high temperatures (600 °C, 900 °C)—were compared using electron backscatter diffraction. The distribution of Σ3 boundaries was compared, and it was found that SABs have a higher fraction of Σ3 boundaries than PABs. Moreover, it was confirmed that the ratio of maximum grain size (area) to the total area of the molten mark in SABs is larger than that in PABs. Thus, reliable discriminant factors were suggested, such as the fraction of Σ3 boundaries and normalized maximum grain size, which can distinguish PABs and SABs. The four discriminant factors, such as the (001)//LD, GAR, fraction of Σ3 boundaries, and fraction of maximum grain size to the total molten-mark area, were verified using the machine learning of t-SNE and Pearson correlation analyses.
6 December 2025 in Materials Science and Engineering: A
Refractory high-entropy alloys (RHEAs) offer attractive strength and thermal stability for extreme environments, yet their limited room-temperature ductility remains a major design challenge. Machine-learning models can accelerate exploration of multicomponent composition spaces, but their reliability is often compromised by strongly imbalanced materials datasets, in which densely sampled regions dominate training while sparsely represented regions may contain valuable high-performance candidates. We propose a distribution-aware, uncertainty-quantified machine-learning framework for predicting elastic properties and guiding ductility-oriented design of RHEAs. A dataset of 3,591 alloys generated from first-principles calculations was described using elemental and thermodynamic descriptors. Physically meaningful extreme observations were retained, and random, density-aware, and hybrid sampling strategies were systematically compared to reduce bias arising from non-uniform target distributions. Bayesian-optimized XGBoost models were trained for eight elastic properties, while 1,000 bootstrap models were used to quantify epistemic uncertainty in both predictions and SHAP-based feature importance. For Pugh’s ratio, 96.5% of test observations were captured within the calibrated 95% prediction interval, and key contributions from mixing entropy, valence-electron-concentration, and electronegativity descriptors were identified. The framework was experimentally validated using two compositionally related alloys with contrasting predicted Pugh’s ratios. Compression testing reproduced the predicted ductility ranking, with the high-Pugh’s-ratio alloy exhibiting more than 15% plastic strain at approximately 1,858 MPa yield strength, demonstrating the practical value of uncertainty-aware materials design despite solidification-induced phase separation and complex dual-phase microstructural evolution.