Inverse Design Of Complex Metasurface Based On Deep Learning

Metamaterial is an artificial synthetic material whose characteristic size is far smaller than the working wavelength(that is,sub-wavelength).It has some special properties that natural materials do not have,such as negative refractive index,negative dielectric constant and negative permeability.Metasurface is a two-dimensional version of metamaterials.Because it still has special electromagnetic properties and easy processing,metasurface design research has become a research hotspot in academia and industry.At present,traditional metasurface design methods mainly rely on empirical reasoning or experimental trial and error to continuously iterate and optimize the material parameters and geometry of metasurfaces to obtain specific metasurfaces with target electromagnetic responses.Because this method involves a large number of full-wave numerical simulations to solve Maxwell’s equations,its efficiency is low.In recent years,with the rapid development of artificial intelligence technology,many scholars have applied deep learning technology to metasurface design to accelerate its design.However,at present,the degree of freedom of metasurface structural units(i.e.meta-atoms)using such methods for inverse design is generally not high.The low degree of freedom of metasurface structure limits the performance of metasurfaces,which contradicts the increasing complexity of metasurface structures,processes and functions in recent years.Therefore,it is urgent to study a new method that can realize the inverse design of complex structural metasurfaces,which will play an important role in the future development of metasurface fields.Based on deep learning technology,combined with genetic algorithms and basic metasurface theoretical knowledge,this paper studies the inverse design of metasurfaces of complex structures with arbitrary symmetric shapes as the design object:1.Aiming at the large number of one-to-many mapping problems in the inverse design of complex structural metasurfaces,this paper proposes a inverse design method of complex structural metasurfaces based on deep learning and genetic algorithm.Firstly,a forward prediction neural network is constructed,which uses the geometric information and material parameter information of the meta-atomic structure to realize the forward prediction of the optical response of metasurfaces of complex structures.Secondly,the trained forward prediction network is combined with genetic algorithm to realize the inverse design of complex structural metasurfaces,and successfully avoids the problem that the neural network is difficult to converge due to the existence of one-to-many mapping problems when directly using neural networks for inverse design.However,because this method is optimized in the whole image domain,it is difficult to obtain regular metasurface element structure in inverse design.2.To address the problem that it is difficult to obtain a regular structure for the inverse design of the complex structure metasurface in the previous design,this paper proposes two methods for the inverse design on-demand generation with the desired structure as a simple regular structure and a complex regular structure as an example.The central idea of both methods is to constrain the solution space,the former restricting the solution space to the domain of simple regular structures by parametric encoding of the structure,and the latter by constructing a variational autoencoder(VAE)neural network so that it learns the geometrical features of complex regular structures and compresses them into the hidden space to achieve constraint on the solution space.At this point,the genetic algorithm only needs to perform an optimization search in the constrained solution space to ensure that the metasurface structure obtained by the inverse design is in accordance with the respective expectation(regular structure)in order to achieve on-demand generation of the inverse design.Finally,We have successively verified on the test set and on custom data,and the results are good.In this paper,we propose a deep learning and genetic algorithm-based inverse design method for complex metasurfaces,and innovatively combine the trained forward predictive neural network,VAE model and genetic algorithm to successfully achieve the inverse design of complex structural metasurfaces and solve the problem of one-to-many mapping and the difficulty of on-demand generation.It is worth mentioning that the essence of the on-demand generation method for inverse design in this paper is to add "constraints" to the optimization algorithm(in this paper,the "constraints" are: to obtain a regular-shaped structure),and different "constraints" can give the method different characteristics,so the method will help to use the optimization algorithm better in any other field.In addition,the method proposed in this paper can be applied not only to the inverse design of meta-atoms but also to other electromagnetic devices,such as antennas,integrated optical circuit devices,etc.