Deep learning-based prediction of structural parameters in FDTD-simulated plasmonic nanostructures

The research creates a new approach to estimate essential dimensions of plasmonic nanoparticles that use the Finite-Difference Time-Domain (FDTD) simulation program. The research team uses EfficientNetB0 alongside ResNet50 and VGG16 deep learning models to obtain quick and exact simulations parameter predictions from simulation image data. The developed dataset consists of dielectric and magnetic field images that stem from FDTD simulated fields through representative materials MgF₂, Au, and glass. The preparation process for the dataset includes a systematic variation of 38 structural parameters for achieving sufficient coverage of potential configurations. VGG16 proved to be the most effective model from the testing group because it attained a training loss 0.1592, validation loss of 0.1607, and test loss 0.1625. The outstanding result shows deep learning techniques can be effectively used to boost nanophotonic device design speeds and optimization processes. The methodology developed in this work has the potential to reduce substantially the computational expenses together with simulation duration for nanostructure engineering processes.