Towards inverse nanophotonic design by predicting plasmonic responses in nanoparticle assemblies with deep learning

Nanoscale structures designed with desired optical responses is a long sought-after goal of the photonics and materials science communities. Much progress has been made recently towards the inverse design challenge, particularly with several deep learning strategies, which was outlined in excellent detail¹. Most of these routes, however, have been focused toward macroscopic or effective optical responses, while nanoscale spatial behavior has been overlooked. Emergent quantum technologies will heavily rely on optical effects at the nanoscale, thus appropriate design choice of nanophotonic elements in this regime is of critical importance. We develop here a deep learning strategy utilizing an encoder-decoder scheme applied to scanning transmission electron microscopy (STEM) monochromated electron energy loss spectroscopy (EELS) data. The nanoscale spatial resolution provided by the electron probe in EELS allows to decode the nanoscale design space, and together with the autoencoder networks, the correlative relationship between plasmonic spectra and geometry is established, ultimately allowing geometry prediction given spectral input (inverse design).

1. Ma, W. et al. Nature Photonics 1–14 (2020)