Research On The Design Of Metasurface With Bound States In The Continuum Based On Machine Learning And Optimization Algorithms

By designing the right shape and arrangement,metasurface can have unique properties.Compared with traditional devices,metasurface devices can regulate the phase and polarization of electromagnetic wave in different wavelengths more precisely.More importantly,the metasurface with bound state in the continuum has the potential to generate ultra-high quality factors,which has attracted people’s interest.However,for complex optical properties,the designer needs a deep theoretical foundation and a large number of calculations,as well as continuous trial and error verification.To simplify the design workload and improve the accuracy,this paper uses machine learning and optimization algorithms for the design,and takes a double-bar metasurface as an example:In terms of machine learning,three networks are built.For the forward prediction problem of the structural parameters to transmission spectra,a forward prediction network is built,and the loss of the network is reduced to 0.0003 by using six hidden layers,relu activation function,Adam optimizer,and dynamic learning rate.For the inverse design problem of transmission spectra to structural parameters,the first method is to use a tandem network,which concatenates the inverse design network on the pre-trained forward prediction network and can avoid the training anomaly problem due to uniqueness,and eventually the loss of the tandem network is reduced to 0.0009.The second method is to use CGAN network.Considering the fact that the same transmission spectrum may be caused by multiple sets of structural parameters,and in order to solve the problem that CGAN network is difficult to train,this paper combines CGAN network and tandem network.The whole network is trained by CGAN network when the number of training is odd,and trained by tandem network when the number of training is even.Finally,the generative mode’s loss is reduced to 0.7336,and the discriminative model’s loss is reduced to 1.3085.Additionally,the tandem network’s loss is reduced to 0.0017.In terms of optimization algorithms,particle swarm optimization and genetic algorithms are used for inverse design.To solve the problem of time-consuming numerical simulations leading to a long running time of the optimization algorithm,a forward prediction network is used instead of numerical simulations,and the running time of 30 hours was replaced by just 133 seconds.For the given fitness function,the MAE between transmission spectrum gotten by using particle swarm optimization and original transmission spectrum is 0.0125.By using genetic algorithms,the MAE is 0.0135.After designing a specific wavelength adaptation function,both optimization algorithms can find the structural parameters corresponding to the transmission spectra with the valley targeted at a specific wavelength.The final results show that both the above neural network model and the optimization algorithm are capable of achieving the design optimization requirements for metasurface devices and thus can be used as an optimization tool for the design of metasurface devices.