(509dj) Interpreting the Optical Absorption Spectra of Gold Nanoclusters By Convolutional Neural Network

UV-Vis absorption spectrometry has been widely utilized in identifying the compositions of metal nanoclusters (NCs) by comparing the experimental spectra with the reference data. However, the application of such method is limited when the optical absorption peaks are difficult to identify (i.e., featureless spectra), most of the time due to the sample being a mixture of many species. Here we develop a machine-learning-based method to interpret the compositions of metal NCs behind the featureless spectra. By implementing a one-dimensional Convolutional Neural Network (CNN), the correlation built between the UV-Vis absorption spectra and the compositions of metal NCs shows much lower mean absolute error value than human can achieve, which can also be verified by good matches between prediction results and experimental results. 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.

From the UV-Vis absorption spectra and ESI mass spectra of 227 Au NC samples, the MAE of the ML model reaches as low as 0.0056. Furthermore, the high accuracy is also demonstrated by several examples when we use experimental techniques to verify the predicted compositions. Through the data-driven training process, chemistry insights such as the size range of Au NCs are gained, suggesting the ML process is not simply searching the possible compositions to match the UV-Vis absorption spectra but featuring the process of finding the most reasonable composition. Additionally, through the correlation between the UV-Vis absorption spectra and compositions of Au NCs, the prediction of UV-Vis absorption spectra can also be conducted. It also shows high accuracies when the predicted spectra are verified by experimental spectra. The ML model developed in this study is a demonstration of elucidating the chemical compositions from the physical properties. It opens a new platform for the identification of metal NC species at molecular precisions (especially for those mass spectrometry and X-ray crystallography cannot identify) and facilitates their further studies in nanochemistry. Last but not least, it makes the high-throughput characterizations in nanochemistry possible owing to the easy-to-use nature of UV-Vis absorption spectrometry.