Inverse design of octagonal plasmonic structure for switching using deep learning

• Leverages deep learning for inverse design of octagonal plasmonic switches. • Adopts the Taguchi method to address the challenge of data collection. • Predicts transmission spectra with high accuracy using a novel neural network. • Enables rapid design optimization, surpassing traditional numerical methods. • Demonstrates the power of deep learning in computational nanophotonics. All-optical plasmonic switches have emerged as essential components for integration into photonic integrated circuits markedly improving optical signal processing. Traditional design methodologies for complex plasmonic structures in switching applications rely on empirical or trial-and-error approaches. Nonetheless, the advent of deep learning heralds a paradigm shift, offering a computational vanguard that promises to streamline nanophotonic simulations. To the best of our knowledge, the potential of deep learning methods for designing state-of-the-art plasmonic switches has not yet been identified. This article unveils an approach that harnesses deep learning to not only predict spectral patterns but also to facilitate the inverse design of plasmonic switches, thus introducing a new era of computational nanophotonics. Our methodology demonstrates the operational efficacy of switches via the nonlinear Kerr effect by employing octagonal nonlinear plasmonic ring resonators crafted from metal-insulator-metal waveguides. The forward model presented herein significantly surpasses traditional numerical methods in terms of computational efficiency, meticulously evaluating a spectrum of configurational parameters to predict transmission spectra characterized by pronounced dips. Through inverse modeling, our approach accurately predicts the design parameters aligned with the desired transmission spectra, offering rapid estimations that starkly contrast with the protracted timelines associated with conventional optimization techniques. The mean squared error was minimal for both the forward and inverse models, recording values of 0.026 and 0.015, respectively. These low losses emphatically affirm the prowess of deep learning in both the forward and inverse engineering of plasmonic switches, setting a benchmark for future explorations in the field.