Feature-based machine learning for the efficient design of nanophotonic structures

Electromagnetic solvers are of paramount importance to simulate the behavior of nanophotonic structures. These solvers can be computationally expensive and therefore their use for design tasks such as optimization can result unfeasible. Machine learning techniques have been proposed in the literature for the modeling and design of photonic devices. In this modeling area, deep neural networks have attracted a lot of attention recently. How-ever, a large number of data samples can be needed to train and validate deep neural networks and therefore the potential speed-up in a design flow offered by these models can be drastically reduced.In this paper, a feature-based machine learning technique is proposed for the design of nanophotonic struc-tures. An adaptive wavelength sampling allows reducing the computational cost of the electromagnetic simu-lations needed to collect the data samples to train and validate the neural networks that model design features as a function of the design parameters. Design features represent very valuable and meaningful information for designers. The complexity of the feature-based neural networks is significantly reduced with respect to neural networks that model the full electromagnetic response sampled over a grid of wavelength values as a function of the design parameters. Also, a significantly reduced number of data samples is achieved in the case of feature -based modeling. The proposed modeling technique is validated by pertinent numerical results related to a multiband absorber nanostructure.