Reconfigurable Magneto-Optoelectronic Devices for Multidimensional Optical Neural Network

Optical neural networks (ONNs) with extremely low latency, low power consumption, and high parallelism, provide an advantageous computational paradigm to address the rapid development of artificial intelligence. Conventional ONNs focus on handling limited information dimensions such as optical amplitude and phase, which are confined to simple small-size image classification, raising demand for significant reconfigurability on the perception of inherent high-dimensional light information. Herein, magneto-optoelectronic devices with polarization sensitivity are theoretically proposed to construct the ONN with high-performance multidimensional recognition, which is composed of 2D magnetic half-metal FeCl₂ and 2H-WSe₂. Polarization sensitivity with photogalvanic effect originates from the space-inversion symmetrical breaking of 2H-WSe₂, yielding the multidimensional perception under zero power consumption. The switchable magnetic configuration of two FeCl₂ contacts with unique half-metal band structures nonvolatilely modulates the amplitude and polarity of photoresponse across the wavelength from ultraviolet to near-infrared. By leveraging multidimensional light encoding, the proposed ONN architecture conducts negative value and nonlinear computations in polarization domain through highly reconfigurable magneto-optoelectronic mechanisms, which achieves up to 93.5% accuracy across complex tasks including 3D object classification, time-series recognition, etc. This work illuminates the potential of magneto-electronics, which extends the applications of ONNs in the real world.