Umeå University's logo

 Nanophotonic and plasmonic research have seen many breakthroughs lately which has created a demand for automated design algorithms to optimize optical elements at the nanometer scale. This work focuses on plasmonic nanostructures that are small metal particles interacting with electromagnetic radiation on length scales typically less than the wavelength. Plasmonic effects from nanostructures are used for enhancing and manipulating electromagnetic fields at the nanometer scale which have seen many applications in sensing requiring an ultra-high sensitivity and a small resolution. This work is about how deep learning methods can be used for designing plasmonic gold nanostructures. In particular, we investigate how convolutional neural networks can be used to predict the optical properties of nanostructures and how conditional generative adversarial networks (cGAN) can be used for designing structures with desired optical properties. The data in this work consist of images with differently shaped nanostructures and the corresponding spectral data for the scattering cross section, the absorption cross section, the polarization rotation and the polarization ellipticity. Utilizing the convolutional EfficientNet architectures, we train a forward model to predict the spectral data of anisotropically shaped nanostructures with images of the structure shape as input. We achieve a mean squared error of 2.5 × 10⁻⁴, 2.5 ×10⁻⁴, 6.0 ×10⁻⁴, and 4.9 ×10⁻⁴ respectively for each variable which agrees with the literature for similar problems. For the inverse design models, we show that label projection can be used to improve the results of two common GAN architectures in combination with a novel label embedding network. We use the Wasserstein GAN method with gradient penalty for training the models to generate images of nanostructure shapes conditioned on spectral data. The best model achieves a pixelwise mean absolute error of 4.9×10⁻³ and an estimated spectral mean absolute error of 8.4×10⁻³ between original and generated images when trained on a dataset containing cylindrical dimer structures. Furthermore, we have shown that the pixelwise mean absolute error is reduced when more conditional input variables are added to the model and that the model can learn different nanostructure shapes when trained on a large dataset containing different anisotropic gold nanostructure shapes. The best pixelwise mean absolute error found is 1.1×10⁻² and the estimated spectral mean absolute error is 1.7 × 10⁻² on the full dataset using all available input data.