Quantitative Analysis of the UV–Vis Spectra for Gold Nanoparticles Powered by Supervised Machine Learning

Surface plasmon resonance is sensitive to the size and shape of gold nanoparticles. The quantitative analysis of the ultraviolet–visible spectra provides information about the structural parameters of the nanoparticles. This task is related to the inverse design problem where machine learning (ML) algorithms show superior performance over classical approaches for problems with many degrees of freedom. If a ML algorithm is used as a black box, it often fails when target experimental data have systematic differences with the theoretical training data set. Our work aims to assess the uncertainties in the structural analysis of gold nanoparticles performed using optical spectroscopy. Therefore, ML is trained over a theoretical data set and then used as a tool to predict the spectrum for any combination of structural parameters. The region of a feasible solution is analyzed via L2 norm contour plots, and the method is extended to multicomponent mixtures where Gaussian distribution mimics the particle size distribution. We also demonstrate that the ML algorithm is able to select only informative features of the spectrum (descriptors) and establish an analytical relation between descriptors of spectra and structural parameters. This work extends the capabilities of optical spectroscopy as an analytical tool for noble metal nanoparticles.