Machine learning-accelerated exploration of ferromagnetic-ferroelectric multiferroics in element substituted 2D transition metal dichalcogenides
Phys. Rev. Materials 9, 124417 – Published 30 December, 2025
DOI: https://doi.org/10.1103/r417-h8nc
Abstract
Two-dimensional multiferroics, especially for ferromagnetic–ferroelectric materials, offer pivotal candidates for advancing fields such as nonvolatile information storage and spintronic devices. However, the mutual competition mechanism between ferroelectricity and magnetism leads to the scarcity of 2D multiferroics, severely limiting their practical deployment. Element substitution was considered as a viable engineering to bring in multiferroicity by tuning the electronic structures, but the complexity of substitution sites and the trial-and-error” approach result in low efficiency for this approach. Herein, we present a machine learning-assisted element substitution strategy to efficiently explore multiferroics in element substituted 2D transition metal dichalcogenides (TMD). By graph neural networks trained on a spin exchange interaction graph dataset, we developed a magnetism model, achieving rapid magnetism prediction with 81% cross-validation accuracy. Using this model to predict and screen the magnetism in the element substituted ferroelectric phase d1T- of 2D TMD, we identify five candidates d1T- ( = Ni, Fe, Zr, Cr; = S, Te) in which substitution-induced magnetism can couple with ferroelectricity to form 2D multiferroics. Taking the d1T--Fe system, we expound its magnetic mechanisms and magnetic tunnel junction, revealing that the origin of magnetism lies in the orbitals of off-centered Fe ions and possessing nonvolatile magnetoelectricity applications. This work provides a robust machine-learning-assisted method for predicting magnetism and paves a way for more efficient discovery of transition-metal-substituted 2D multiferroics.
