A multi-objective permittivity optimization for object classification at the speed of light

This paper presents a numerical demonstration of the use of the adjoint method for permittivity optimization to design a dielectric medium capable of object classification at the speed of light. In a two-dimensional setup, the system comprises an input waveguide, a design region, and three output ports made of a lossless dielectric material. The design medium is optimized to guide light into specific output ports based on the type and variation of scatterers placed between the input waveguide and the design region. For proof of concept, scatterers derived from the MNIST dataset's digits 0, 1, and 2 are used to represent different object classes with varying shapes and sizes. The optimization process dynamically adjusts the material distribution within the design region to maximize classification performance. The final structure achieved a classification accuracy of 96.3%, with light successfully directed to the correct output port corresponding to each scatterer class. This work demonstrates the potential of permittivity optimization for developing advanced photonic devices capable of ultrafast object recognition, paving the way for future research in three-dimensional designs and more complex classification tasks.