Abstract:Artificially designed photonics devices have promising applications in various fields of modern optics. The design of conventional photonics devices is usually based on a known physical model,and then the structure is optimized by numerical simulation methods. Since the device structure relies heavily on the a priori model,the degree of freedom of the conventional optimized design is limited. In recent years,with the increasing demand for high- performance photonic devices,inverse design methods with higher design degrees of freedom have been rapidly developed. Currently,the most widely used inverse design method is the gradient descent algorithm, which can achieve fast iterative approximation of the target by using the gradient information of the objective function on the variables. For problems where the gradient is difficult to solve or uncertain,genetic algorithms or particle swarm algorithms can generally be used, which find the global optimal solution by simulating the evolutionary process of organisms and foraging of populations,respectively,and thus do not require gradient information. In recent years,with the rapid development of artificial intelligence,neural network-based machine learning algorithms have attracted widespread attention in various scientific fields. Neural network algorithms are flexible in regulation and can be combined with a variety of algorithms,but the models lack universality and require corresponding data sets for different physical models. It can be seen that different inverse design methods have different advantages and limitations, so for different design problems, the physical model needs to be evaluated and a suitable inverse design method needs to be selected. Compared with traditional parametric design methods,inverse design methods can yield more complex and diverse device structures with superior performance. In addition to using a single inverse design method,the combination of multiple methods is also beneficial to improve the computational efficiency. For example,combining deep learning with genetic algorithms not only improves the computational speed of genetic algorithms,but also makes use of the gradient- free feature of genetic algorithms to find the global optimal solution. The inverse design method breaks the design limitations of traditional methods and can achieve efficient parameter optimization in the full parameter space,thus making it easier to obtain device structures with extremely high performance. This paper summarizes the common methods for inverse design of optoelectronic devices and gives specific applications of inverse design in various optoelectronic fields. With the continuous development of computer science,the inverse design of photonic devices has shown unparalleled potential. Compared with traditional design methods,intelligent inverse design methods are more efficient and offer greater freedom,providing new solutions for achieving high-performance photonic devices. In various fields of photonics,the inverse design approach allows a higher degree of freedom in optical field modulation and enables the design of various high-performance photonic devices from a demand perspective.

Selecting robust silicon photonic designs after Bayesian optimization without extra simulations

Neural Network-Based Surrogate Model for Inverse Design of Metasurfaces

Inverse Design for Silicon Photonics: From Iterative Optimization Algorithms to Deep Neural Networks

BOSON$^{-1}$: Understanding and Enabling Physically-Robust Photonic Inverse Design with Adaptive Variation-Aware Subspace Optimization

Automatic Selection of Structure Parameters of Silicon on Insulator Lateral Power Device Using Bayesian Optimization

Stochastic and multi-objective design of photonic devices with machine learning

Design and Electro-Optical Co-Optimization of High-Speed Travelling Wave Silicon-Organic Hybrid Modulators

Design of high-performance entangling logic in silicon quantum dot systems with Bayesian optimization

Fabrication uncertainty guided design optimization of a photonic crystal cavity by using Gaussian processes

Improved Inverse Design of Polarization Splitter with Advanced Bayesian Optimization

Advances in machine learning optimization for classical and quantum photonics

A Direct Coupled Solution Methodology for Efficient Robust Optimizations of Inverse Problems under Uncertainty

An Efficient Direct Search Methodology for Robust Optimization of Electromagnetic Devices

Bayesian Optimization of Photonic Nanojets Generated by Multilayer Dielectric Structures

Optimizing Photonic Nanostructures via Multi-fidelity Gaussian Processes

Advances of Inverse Design in Photonics(Invited)

Benchmarking five global optimization approaches for nano-optical shape optimization and parameter reconstruction

Accelerated Design of Microring Lasers with Multi-Objective Bayesian Optimization

Integrated silicon photonic device design by attractor selection mechanism based on artificial neural networks: optical coupler and asymmetric light transmitter.

Optimization design of steady optical pin beam using genetic algorithm

Mastering Silicon Photonics Device Design for Scalable and Robust Optical Neural Networks