CENTER FOR THE SCIENCE OF
SYNTHESIS ACROSS SCALES
Spectra-to-Structure Prediction of Nanostructure Geometries
​
September 16, 2025
Scientific Achievement
-
​​Dual-variational autoencoders (dual-VAEs) trained on sparse datasets predict the geometries of anisotropic nanostructures with effectiveness approaching conventional physics-based models
​​
​Significance and Impact
-
Spectra-to-structure predictions enable sub-diffraction limit structural determinations
The work establishes dual-VAEs as a generic model for inverse nano-optical problems involving both structural and far-field constraints​​
​
Research Details
-
We collect linearly polarized dark-field scattering spectra of plasmonic Au nanorods and train both spectral and imaging data in shared latent space with enforced similarity to enable spectra-to-structure prediction
Orientational angle and aspect ratio determination of Au nanorods (NRs) is achieved via dual-variational autoencoders (dual-VAEs)
Sun, M.; Huang, Z.; Yaman, M. Y.; Ziatdinov, M.; Kalinin, S. V.; Ginger, D. S. (2025) J. Phys. Chem. C 129:22066-22074. DOI:10.1021/acs.jpcc.5c07879
​
Work performed at the University of Washington and University of Tennessee Knoxville.