Identification of chemical compositions from “featureless” optical absorption spectra: Machine learning predictions and experimental validations

Rapid and accurate chemical composition identification is critically important in chemistry. While it can be achieved with optical absorption spectrometry by comparing the experimental spectra with the reference data when the chemical compositions are simple, such application is limited in more complicated scenarios especially in nano-scale research. This is due to the difficulties in identifying optical absorption peaks (i.e., from “featureless” spectra) arose from the complexity. In this work, using the ultraviolet—visible (UV—Vis) absorption spectra of metal nanoclusters (NCs) as a demonstration, we develop a machine-learning-based method to unravel the compositions of metal NCs behind the “featureless” spectra. By implementing a one-dimensional convolutional neural network, good matches between prediction results and experimental results and low mean absolute error values are achieved on these optical absorption spectra that human cannot interpret. This work opens a door for the identification of nanomaterials at molecular precision from their optical properties, paving the way to rapid and high-throughput characterizations.