Machine Learning Study of the Magnetic Ordering in 2D Materials

Abstract

Magnetic materials have been applied in a large variety of technologies, from data storage to quantum devices. The development of two-dimensional (2D) materials has opened new arenas for magnetic compounds, even when classical theories discourage their examination. Here we propose a machinelearning-based strategy to predict and understand magnetic ordering in 2D materials. This strategy couples the prediction of the existence of magnetism in 2D materials using a random forest and the Shapley additive explanations method with material maps defined by atomic features predicting the magnetic ordering (ferromagnetic or antiferromagnetic). While the random forest model predicts magnetism with an accuracy of 86%, the material maps obtained by the sure independence screening and sparsifying method have an accuracy of ∼90% in predicting the magnetic ordering. Our model indicates that 3d transition metals, halides, and structural clusters with regular transition-metal sublattices have a positive contribution in the total weight deciding the existence of magnetism in 2D compounds. This behavior is associated with the competition between crystal field and exchange splitting. The machine learning model also indicates that the atomic spin orbit coupling (SOC) is a determinant feature for the identification of the patterns separating ferro-from antiferromagnetic order. The proposed strategy is used to identify novel 2D magnetic compounds that, together with the fundamental trends in the chemical and structural space, pave novel routes for experimental exploration.

Multifunctional 2D g-C4N3/MoS2 vdW Heterostructure-Based Nanodevices: Spin Filtering and Gas Sensing Properties

“…Ultrathin films of two-dimensional (2D) magnetic layered materials and the integrated van der Waals (vdW) planar heterostructures have attracted great interest due to their attractive physical properties, which provide new opportunities for the development of magnetic nanodevices. − Utilizing the coupling advantages of bilayer or multilayer 2D materials, vdW heterojunctions have flourished in a wide range of physical applications, including semiconductor field-effect transistors, , atom-thin superconductors, optoelectronics devices, and low-dimensional switching devices . For the low-dimension materials, different degrees of freedom of spin can exhibit exotic physical phenomena and regulate their physical properties over a wide range. − Through improvements in electronic spin injection and manipulation techniques, plenty of 2D vdW heterostructures seem to be promising for spintronics applications through improvements in electronic spin injection and manipulation techniques .…”

Machine learning magnetism classifiers from atomic coordinates

“…The necessity to include atomic structures for prediction can be argued by the facts that some materials are described by the same chemical formulas but have different atomic structures, thus completely different magnetic structures. To further elaborate, it is hard to compare the performance of our model with some prior works that did not encode crystal structures or used other descriptors to encode, because the training data are different (most prior works focused on 2D materials ( Nelson and Sanvito, 2019 ; Bassman et al., 2018 ; Xie et al., 2021 )), and the prediction outputs are different(some prior works predict critical temperatures of ferromagnetic ( Nelson and Sanvito, 2019 ) or superconducting ( Bassman et al., 2018 ; Konno et al., 2021 ), or classify magnetic orders but only between two classes FM and AFM ( Acosta et al., 2022 )). In order to make a comparison and show the necessity to include atomic structures, we performed some comparable models, with only chemical composition as inputs.…”

Introduzindo aprendizado de máquina em cursos de física: o caso do rolamento no plano inclinado

“…Pelo contraste entre os modelos teóricos e de aprendizado de máquina, um pesquisador poderia concluir que as mesmas variáveis físicas selecionadas podem ser usadas para gerar um modelo mais acurado, o que é um teste indicativo sobre a validade das hipóteses assumidas no modelo físico -mas que exige cuidados com relação as instâncias de dados usadas para treinamento e teste (cuja separação pode gerar certo vazamento de informação entre as etapas, emulando mais informação do que o modelo deveria ter acesso). De uma perspectiva diferente, pesquisas que envolvam a análise de muitos dados (que podem ser rotulados por humanos em quantidade muito menor do que por um algoritmo automatizado) podem se valer dessas técnicas, por exemplo: a detecção de exoplanetas, [30]; a predição de ordem magnética em materiais bidimensionais [31]; e a análise de experimentos de altas energias [32,33]. Em qualquer aplicação, como ferramentas de aprendizado automatizado, os algoritmos também podem falhar em suas previsões, seja devido ao sub-ajuste ou ao sobre-ajuste, ou pela própria natureza estatística de seu funcionamento.…”