Due to their excellent magnetic properties, the sintered Neodymium-Iron-Boron (NdFeB) magnets are currently the widely used permanent magnets. These magnets constitute rare earth elements (REE) that are scarcely available and have adverse effects on the environment due to mining and processing, leading to the limited and controlled supply of REEs into the markets and the rise in cost. Therefore, searching for novel magnetic phases that require less REE or less expensive REEs is essential. The shift in the paradigm for materials development starts from the pure empirical methods followed by a theoretical-based approach and a computational science approach involving simulations such as the density functional theory (DFT), respectively. These paradigms have led to the generation of a vast amount of data along with advancements in computers leading to the emergence of data-driven approaches that can generate both forward and inverse models to find a better relationship in processing, structure, properties, and performance (PSPP). Forward models have been used to study the cause and effects of the processing, composition, and structure on the material property. On the other hand, inverse models have been used to design and optimize materials with application goals.

In this paper, we present a supervised machine learning (ML) based weighted voting regression model for the prediction of the Curie temperature (T_c) and saturation magnetization (M_s) or saturation polarization (J_s = µ₀M_s) for the magnetic 14:2:1 phase family, trained using the chemical compositional-based features. The developed model is a single generalized model that applies to the 14:2:1 phase with a ternary to a senary system that alloys different rare earth and transition metals. The dataset consists of chemical compositional-based features covering 33 elements. The reference values for performance evaluation are the reported experimental T_c [K] and M_s [µ_B/f.u.] values measured at room temperature across literature sources. The effectiveness of the trained ML model on the 14:2:1 phase containing heavy rare earth, light rare earth, and a combination of both has also been evaluated. The trained model has a low mean absolute error of 16 K for T_c prediction and an absolute prediction error of less than 2 % for M_s on the unseen test dataset. In figure 1, the results for Ms prediction are shown. Such data-driven models can be the basis for developing magnets with desired properties.