Predictability of Localized Plasmonic Responses in Nanoparticle Assemblies

Abstract

Design of nanoscale structures with desired optical properties is a key task for nanophotonics. Here, the correlative relationship between local nanoparticle geometries and their plasmonic responses is established using encoder‐decoder neural networks. In the im2spec network, the relationship between local particle geometries and local spectra is established via encoding the observed geometries to a small number of latent variables and subsequently decoding into plasmonic spectra; in the spec2im network, the relationship is reversed. Surprisingly, these reduced descriptions allow high‐veracity predictions of local responses based on geometries for fixed compositions and surface chemical states. Analysis of the latent space distributions and the corresponding decoded and closest (in latent space) encoded images yields insight into the generative mechanisms of plasmonic interactions in the nanoparticle arrays. Ultimately, this approach creates a path toward determining configurations that yield the spectrum closest to the desired one, paving the way for stochastic design of nanoplasmonic structures.

Machine learning in scanning transmission electron microscopy

“…The autoencoder concept can be extended towards learning correlative relationships between structure in an image and property in spectral data as has been demonstrated with the im2spec encoder-decoder models. 209 Finally, transformation-invariant variational autoencoders (VAE) build upon classical autoencoders by making the reconstruction process probabilistic and incorporating prior knowledge into the latent space structure 210 . Figure 4 shows the application of rotationally invariant VAE to the analysis of graphene data.…”

Instantaneous Property Prediction and Inverse Design of Plasmonic Nanostructures Using Machine Learning: Current Applications and Future Directions

“…It was concluded that the encoder-decoder network with a high dimensional latent space (10D) was able to yield better accuracy compared to a 2D latent space. Noticeably, the use of latent space allowed transferable ML cognition that could be used by other data sets [90].…”

“…The bidirectional network was made possible by a type of representation learning method: the auto-encoding neural network. The network encoder, through multiple non-linear layers of abstraction, represents the input data in the form of latent space, which then is transformed to the output data by the decoder [90]. Besides the DNN powered data representation, He et al further simplified the near field (electric-field enhancements) response by cherry-picking and downsizing the collected data.…”

“…Sub-figures i-iv correspond to different locations of the nanoparticle assembly. Adapted with permission from Ref [90]…”

Autonomous Experiments in Scanning Probe Microscopy and Spectroscopy: Choosing Where to Explore Polarization Dynamics in Ferroelectrics

“…For instance, a subset of the current authors have previously shown that autoencoders can be utilized to learn the mapping between imaging and spectral response. 39,40 These studies, in addition to basic physics considerations, would suggest that incorporating the knowledge of the domain structure should be highly beneficial in optimizing the sampling process. To improve on the process, we utilized a secondary approach where, in addition to utilizing the measured spectral values in the GP surrogate model, we also utilize the high-resolution imaging data itself as a secondary input channel, to aid predictions.…”