Solve And Design Optical Inverse Problems Using Neural Networks

The inverse design of nanophotonic devices is one of the important research topics in nanophotonics.The typical inverse design problem achieves the desired performance through electromagnetic simulation and inverse design methods.However,with the development of technology,the structure of nanophotonic devices is more and more complex.For these more complex structures and higher dimensional optimization space,a fast forward simulation and inverse design method will accelerate the device design process effectively.Therefore,this paper proposes an inverse design method based on multilayer perceptron(MLP)model.These trained MLP models can replace the expensive calculations of electromagnetic simulation.As for the inverse design method,this paper is mainly divided into three parts:In the first part,this paper proposes a hybrid optimization algorithm combining differential evolution algorithm and gradient optimization algorithm,hereinafter referred to as automatic optimization gradient optimization algorithm(AOGOA),in order to alleviate the sensitivity of the initial values for the gradient optimization algorithm and the slow convergence speed of the population-based optimization algorithm.The AOGOA first searches in the global parameter space depending on the idea of difference iteration in order to find an initial design point which is beneficial to the gradient optimization algorithm for further search,and then uses the gradient optimization algorithm for local search.The results show that this algorithm can effectively accelerate the convergence process of the whole inverse design.In the second part,a new gradient optimization strategy based on MLP model is proposed.For the nanophotonic inverse design.There are two types of parameters that can be optimized:discrete parameters and continuous parameters.In order to optimize discrete and continuous optical parameters simultaneously,an unsupervised inverse design neural network and a new constraint function are used to optimize the discrete parameters such as material and period number of photonic crystals,while the continuous parameters of dielectric layer thickness are directly updated by the gradient optimization algorithm.The last part is based on the above two parts.In order to optimize discrete and continuous optical parameters simultaneously,and to overcome the problem that gradient optimization algorithm depends on initial value,a global hybrid optimization algorithm combining genetic algorithm and gradient optimization algorithm is proposed.The global hybrid optimization method firstly uses genetic algorithm to search for an appropriate initial solution in the binary parameter space,and then uses gradient optimization algorithm to search for the possible optimal solution in the surrounding continuous parameter space.The algorithm considers both global search and local search,which can accelerate the convergence speed.We applied this method to the design of two different optical scenes,including multilayer nanoparticles and multilayer nanofilms,and the design results show that the algorithm has good robustness.We also implemented an optimal design on a small dataset by applying transfer learning between two different optical scenes.