Inverse Design Of Functional Multimode Interference Coupler Based On Neural Network

In recent years,the development of silicon photonics technology has been relatively mature,and the basic silicon waveguide devices have been deeply studied.The powerful design kit is very important for photonic integrated circuits,so a diverse library of on-chip photonic functional elements has been established.Finely designed integrated optical devices typically come from a good understanding of physics and intuitive attempts or multiple adjustments to several device parameters.However,this design principle is based on conventional device geometries with very limited parameter space.In practice,this method cannot fully explore all the properties possessed by the device,and fully utilizes the semiconductor manufacturing capabilities of the prior art.Research on the application of reverse-based design to photonic integrated circuits was carried out.One attractive device geometry is a digital meta-structure that discretes a dielectric waveguide into a plurality of pixels having binary or multi-level material properties.The material properties of each pixel are determined by a Direct binary search(DBS)algorithm.This artificial non-uniformity allows for more flexible refractive index engineering at subwavelength(SW)scales compared to conventional periodic or photonic bandgap structures.Based on the DBS method,there have been demonstrations of various digital components,such as power dividers with flexible ratios,sharp bends,mode multiplexers,ultra-compact adiabatic cones,grating couplers,etc.,although this design method has been shown There are great potentials for many new nanophotonic devices,but the computational cost of optimization is a big problem.The excellent prediction and fitting ability of neural networks can solve this problem under certain conditions.In this paper,the ultra-wideband multimode interference coupler(MMI)and attenuator are selected as the design devices.The traditional pixel-based optimization method based on digital metamaterials is used to optimize the device by optical simulation software.The design,and the hybrid system using digital metamaterials and neural networks are used to optimize the design of the device,and the neural network reverse optimization design method is used to design the target device.The results are analyzed and compared.Compared with the existing pixel-based optimization design method based on digital metamaterials,the hybrid system design method and the reverse design method shorten the time cost required for the design when the obtained result is the same or similar to the design target value.Through the prediction function of artificial neural network design method and the ability to simulate complex input-output relationship,we accelerate the development cycle of waveguide devices in automatic optimization design.We believe that by continuing to improve the artificial neural network method,the prediction results can be more accurate.And the scope of prediction is more abundant.Moreover,with the continuous improvement of the artificial neural network method,more features of the device can be mined,and performances that are difficult to achieve by conventional devices can be realized.