Unsupervised learning of nanophotonic optical responses via embedded spectra in t-distributed stochastic neighbor embedding methods

Improving and designing new nanophotonic systems require a deep understanding of what structural parameters can be tuned in a given environment to modify their optical properties. Traditionally, this process relied on linear or mean squared error measures which are not always suitable for highly similar spectra. However, recent progress has been made for such dataset by employing graphical networks to learn underlying non-linear structures. Here we apply a novel mapping method of embedded space via injection of additional synthetic spectra to our complex original dataset of Mie scattering spectra and quantify the performance of our approximated model. We also discuss how the underlying latent space can be interpreted as a series of Markov processes which map the resonance shifts.