Forward modeling for metamaterial design using feature-based machine learning

Machine learning techniques have been proposed in the literature for the modeling of photonic devices. In this paper, a modeling technique based on system identification, feature extraction, and machine learning methods is proposed for the design of photonic devices. Design features of interest are extracted based on a system identification step that uses a few samples of the electromagnetic device response. This system identification step allows saving computational resources significantly while collecting the data needed for the further machine learning step. Modeling design features instead of the wavelength-dependent device response as a function of the design parameters allows compacting the output space of interest in neural networks and reducing related model complexity issues. These features can be modeled as a function of design parameters by means of neural networks. The generated neural networks are of very limited complexity. Design features represent very valuable and meaningful information for designers. Numerical results successfully validate the proposed technique.