The need to reduce carbon dioxide emissions has accelerated the demand for green energy technologies such as electric vehicles and wind turbines, which heavily rely on Nd-Fe-B based magnets. However, the limited availability of Dy, Tb and Nd, which are essential components of these magnets, has renewed interest in Fe-rich SmFe12-based compounds, which offer superior intrinsic magnetic properties at elevated temperatures and a reduced rare-earth fraction compared with Nd2Fe14B. To translate these intrinsic magnetic properties into extrinsic ones, an optimal microstructure is needed in which SmFe₁₂ grains are isolated by a low-melting-point intergranular phase. Consequently, phase diagrams are indispensable for identifying the compositional and thermal conditions that enable such coexistence [1].

In this study, we developed an active-learning framework to accelerate the Sm-Fe-V phase development in order to seek this coexistence. The model iteratively selects the most informative experiments, allowing us to refine the Sm-Fe-V phase diagram while minimizing the number of samples synthesized. We began with a “0th” iteration by digitizing existing literature phase-equilibria data [2] and used Bayesian optimization with K-fold cross-validation to tune hyperparameters for SVM, linear classifier, RF, KNN and NN. The two highest-accuracy models (NN + RF) were then combined into an ensemble predictor. At each iteration, the ensemble’s confidence map directs synthesis of four ingots-targeting phase-boundary refinement or high-performance regions, whose measured equilibria update the training set.

Target alloys were produced by induction melting and annealed at 1100 °C for 20 h, then characterized for phases coexistence. New data were given higher weight in subsequent model fits. After four cycles, the 1:12 phase proved stable over a broader V range and its two-phase (1:12 + liquid) field was more confined than previously reported. These findings necessitate revision of existing boundaries and demonstrate that active learning can both accelerate and enhance phase-diagram development in complex materials systems [3].

References

[1] P. Tozman, H. Sepehri-Amin, K. Hono, Scripta Materialia, 2021, 194, 113686.

[2] S. Sugimoto, T. Shimono, H. Nakamura, T. Kagotami, M. Okada, M. Homma, Materials Transactions, 1996, 37, 494-498.

[3] R. Katsube, K. Terayama, R. Tamura, Y. Nose, ACS Materials Letters, 2020, 2, 571-575.