How to Select the Proper Gold Permittivity for Generating Training Data for Deep Neural Networks for Analysis of the Nanoparticle Colloid Size Distributions from Their UV–Vis–NIR Extinction Spectra?

The investigation of colloidal nanoparticle size distribution has emerged as one of the most important yet routinely executed measurements due to its wide range of applications. Deep learning has recently become a valuable tool for spectroscopic data interpretation including localized surface plasmon resonances in the UV–vis–NIR extinction spectra. However, deep neural networks (DNNs) require a large number of training data instances to make accurate predictions. Researchers frequently turn to computational methods, such as Mie theory, for this task. Unfortunately, the influence of permittivity selection on neural network prediction results for plasmonic metal nanoparticles is often underinvestigated or even entirely overlooked. In this article, we thoroughly examined the relation between gold nanoparticle size distributions, their extinction spectra, and permittivities provided by different authors in the scientific literature. We investigated the effects of using numerically generated training data originating from different permittivities on DNN predictions in detail. The relationship between permittivity and extinction spectra for wet-chemistry-synthesized Au nanoparticles was verified, which directly applies to any spherical or near-spherical plasmonic nanoparticles.