Machine Learning and Small Data-Guided Optimization of Silica Shell Morphology on Gold Nanorods

Anisotropic plasmonic nanorods offer a wide range of
applications
in photovoltaics, energy conversion, sensing, and surface-enhanced
Raman spectroscopies. However, achieving control over the size and
shape of the surface overcoating on these nanorods remains a challenge
due to the complexity arising from the multistep wet chemical processes
involved in their experimental synthesis. Here, we show that by employing
data imputation and data augmentation methods, we can minimize the
limitations of a small experimental data set and successfully train
supervised machine learning models that can optimize the experimental
synthesis. Using a small data set collected from 30 multistep syntheses
of silica-overcoated gold nanorods (GNRs) characterized by optical
extinction spectroscopy and transmission electron microscopy, we trained
complementary supervised models to predict the overcoating shape of
the nanorods using optical spectral features. The effects of experimental
parameters and measurements made during different stages of the synthesis
were analyzed. Our approach enabled us to design an experimental synthesis
recipe to yield a target SiO2 overcoating shape on GNRs
employing inverse design optimization. The developed workflow can
be extended to other plasmonic nanoparticles and multistage synthesis
experiments, where a limited data set is available to understand the
effects of synthesis parameters and to establish correlations between
measurements and synthetic yields.