Korean Journal of Materials Research ISSN:1225-0562(Print) 2287-7258(Online) 한국재료학회지

Sintered NdFeB permanent magnets are indispensable in electric-vehicle motors, wind-turbine generators and other applications, and their performance at elevated temperatures is critically governed by coercivity. Although machine learning is widely used to predict such properties, the inability to quantify predictive reliability makes it difficult to assess, a priori, the risk in unexplored compositional regions. In this study, a random-forest regression model was trained on an experimental NdFeB dataset of 262 samples, and the predictive uncertainty was decomposed into epistemic and aleatoric components based on variance conservation. To create an extrapolation scenario, 137 dysprosium (Dy)-free samples formed the training set and 125 Dy-containing samples comprised the test set. The model, accurate on the training set (R² = 0.965), collapsed on the test set (R² = -0.994), while the total predictive uncertainty increased roughly fifteen-fold. When the eleven most promising candidates were selected using an acquisition score derived from the two uncertainties, all clustered in the low-Dy region. However, in a direct validation, adding such a “favorable” low-Dy candidate barely changed test performance (∆R² = +0.029), whereas adding a high-Dy sample—deemed “unfavorable” by the score—substantially improved it (∆R² = +0.508). These findings indicate that the simple variance-subtraction scheme can misclassify model bias as aleatoric uncertainty under extrapolation, and suggest that, from an active-learning perspective, preferentially acquiring samples in high-error extrapolative regions is a more data-efficient strategy.