Akshar Varma1,Xiaoyu Zhang1,Brian Lejeune1,Laura Cebada Almagro2,Rafael Perez del Real3,4,M. Pilar Marin Palacios2,3,O. Fitchorova1,Laura Lewis1,Ravi Sundaram1
Northeastern University1,Complutense University of Madrid2,Unidad Asociada (CSIC)3,Instituto de Ciencia de Materiales de Madrid4
Akshar Varma1,Xiaoyu Zhang1,Brian Lejeune1,Laura Cebada Almagro2,Rafael Perez del Real3,4,M. Pilar Marin Palacios2,3,O. Fitchorova1,Laura Lewis1,Ravi Sundaram1
Northeastern University1,Complutense University of Madrid2,Unidad Asociada (CSIC)3,Instituto de Ciencia de Materiales de Madrid4
A remote authentication system that is secure and high-throughput requires rapid disambiguation of uniquely identifiable signatures. In pursuit of this goal, an auto-encoder-inspired deep neural network architecture is utilized on data describing the high-frequency (GHz) response of a family of unique magnetic materials, glass-coated amorphous magnetic microwires. These ultrasoft ferromagnetic wires exhibit a very high sensitivity to applied magnetic fields that is derived from their special circumferential magnetic anisotropy. In this work, investigation was focused on the response of amorphous magnetic microwires comprised of a CoFeSiB-based cylindrical core surrounded by borosilicate glass (<i>i.e.,</i>pyrex), delivering an outer diameter of approximately 100 microns. The interplay between the atomic and magnetic structure of these microwires, combined with the remanent tensile stress acquired during synthesis, produces interesting magnetic properties and a high-frequency giant magnetoimpedance (GMI) effect that allows them to modulate the scattering of incident GHz-range electromagnetic waves.<br/><br/>Two-dimensional planar arrangements of amorphous magnetic microwires were interrogated to produce a set of complex spectra. These data were utilized to obtain a parametrization of the response function, reducing the challenge of maximizing unique signatures to a geometric problem in parameter space. Recently, computational efforts based on machine learning have successfully reproduced the response distribution of real magnetic data with a mean square error below 0.01 achieved on unseen data. Work is continuing to refine the parameter space and magnetic entity composition, properties and physical arrangements.