The lack of uniqueness poses a common challenge in the inverse design of nanophotonic structures. This issue arises from the presence of multiple sets of design parameters that yield identical output configurations. Several existing machine learning methods offer promising solutions for enabling flexibility in selecting sets of design variables. However, these approaches still face significant challenges like limited output diversity, and training instability. These issues constrain the variety of solutions attainable for a particular target response. To overcome these challenges in the realm of multi-solution inverse design problems in nanophotonics, we have developed a GA-βCVAE architecture based on generative models. This approach effectively tackles such problems, thereby expanding the range of distinct solutions achievable in nanophotonics. Our approach combines a genetic algorithm with a β conditional variational autoencoder neural network based generative model, capturing multiple distinct solutions for a given target. It provides flexibility in selecting design variables. We successfully applied it to design TiO2-SiO2 based multilayer thin films in the visible range (400–750 nm), with layer thickness as the design variable. Experimental verification confirmed the effectiveness of our proposed architecture.
Read full abstract