Gated Recurrent Neural Network for Predicting the\nPlasmonic Colloid Composition from Spectra

In current research on the synthesis\nof colloidal nanostructures,\nthe size and morphology of nanoparticles still exhibit certain dispersion\nand variation from batch to batch. Characterization of size distribution\nand morphology distribution of nanoparticles often requires techniques\nsuch as scanning electron microscopy or transmission electron microscopy,\nwhich involve high vacuum environments, are time-consuming, and costly.\nExperienced researchers can roughly estimate the size and distribution\nof nanostructure from spectra for a given synthetic route, but the\naccuracy is often limited. This paper reports the potential of using\nneural networks to accurately predict the composition of colloidal\nnanostructures from spectra. We address several fundamental issues\nin neural network prediction of colloidal composition. We first demonstrate\nthe prediction of the composition of <i>a colloidal binary mixture\nof gold nanoparticles</i> using a gated recurrent neural network\n(GRU). The evolution of prediction errors for scattering, absorption,\nand extinction spectra of nanostructures with sizes ranging from 5\nto 120 nm are analyzed. Furthermore, we demonstrate that the neural\nnetwork model operates robustly under white noise in experimental\ntesting scenarios. Compared to fully connected neural networks, the\ngated recurrent unit exhibits better testing accuracy in spectral\nprediction. When confronted with experimental data that deviates from\nsimulation outputs, minor adjustments to the training set can allow\nthe predictions to align closely with the experimental spectra, paving\nthe way for the characterization of complex colloidal compositions\nwith artificial intelligence.