Deep learning-driven forward and inverse design of nanophotonic nanohole arrays: streamlining design for tailored...

In nanophotonics, nanohole arrays (NHAs) are periodic arrangements of nanoscale apertures in thin films that provide diverse optical functionalities essential for various applications. Fully studying NHAs' optical properties and optimizing performance demands understanding both materials and geometric parameters, which presents a computational challenge due to numerous potential combinations. Efficient computational modeling is critical for overcoming this challenge and optimizing NHA-based device performance. Traditional approaches rely on time-consuming numerical simulation processes for device design and optimization. However, using a deep learning approach offers an efficient solution for NHAs design. In this work, a deep neural network within the forward modeling framework accurately predicts the optical properties of NHAs by using device structure data such as periodicity and hole radius as model inputs. We also compare three deep learning-based inverse modeling approaches-fully connected neural network, convolutional neural network, and tandem neural network-to provide approximate solutions for NHA structures based on their optical responses. Once trained, the DNN accurately predicts the desired result in milliseconds, enabling repeated use without wasting computational resources. The models are trained using over 6000 samples from a dataset obtained by finite-difference time-domain (FDTD) simulations. The forward model accurately predicts transmission spectra, while the inverse model reliably infers material attributes, lattice geometries, and structural parameters from the spectra. The forward model accurately predicts transmission spectra, with an average Mean Squared Error (MSE) of 2.44 × 10^-4. In most cases, the inverse design demonstrates high accuracy with deviations of less than 1.5 nm for critical geometrical parameters. For experimental verification, gold nanohole arrays are fabricated using deep UV lithography. Validation against experimental data demonstrates the models' robustness and precision. These findings show that the trained DNN models offer accurate predictions about the optical behavior of NHAs.