Machine learning guided design of RE-Fe-B(RE=PrNd,La,Ce) with comprehensive high performance

Abstract

The optimization of high-abundance REFeB (RE=PrNd,La,Ce) permanent magnets has been a significant research focus, but traditional trial-and-error methods are challenging due to high costs and time consumption. Here, we propose a machine-learning approach to accelerate the design of melt-spun high-abundance (PrNd,La,Ce)-Fe-B ribbons based on the database incorporating elemental electronegativity with the composition and magnetic performance collected from literature. By combining heuristic optimization algorithms and ensemble strategies, we developed accurate and robust machine learning models, allowing for rapid evaluation of comprehensive magnetic performance across different compositions in high-dimensional data spaces and discovering high-performance REFeB permanent magnets with high-abundance rare earth elements. Utilizing the established models, by balancing three magnetic properties of coercivity, remanence and maximum magnetic energy product, we discovered a compositional range with optimal overall magnetic performance and high proportions of high-abundance rare earth elements (up to 40 % La and 20 % Ce of the total rare earth content) for the magnets of (PrNd_xLa_yCe_1-x-y)₁₂Fe₈₂B₆, which were verified by experiments with accuracies exceeding 90 %. Within this range, four cost-effective compositions were identified, among which the best composition, (Pr,Nd)_8.1La_3.6Ce_0.3Fe₈₂B₆, achieved a 31.3 % cost reduction while retaining 86.4 % of the magnetic performance. This study advances the optimization of REFeB compositions with high-abundance rare earth elements, demonstrating the enormous potential of machine-learning approach in the design and development of high-performance and cost-effective REFeB permanent magnets.