Machine Learning Captures Synthetic Intuitions for Hollow Nanostructures

Abstract

Machine learning (ML) has been transforming materials science, making great strides in property prediction and materials discovery; however, how ML can bring benefits to nanostructure synthesis, given that the synthesis routines of nanostructures are usually complex with strongly correlated, highly dimensional parameters with narrow experimental window, is rarely explored. Here, we choose hollow nanospheres and emulsion interfacial polymerization as the prototype of material and synthesis routine and propose an ML method to capture synthetic intuitions for guiding the preparation. The quantitative relationship between the reactants, synthesis parameters, and the product morphologies is captured, and the contributions of each feature are analyzed to optimize the reactant amounts and synthesis parameters for hollow and homogeneous spheres. The experimental results are generally consistent with the ML predictions, with the f1-scores of 0.750 and 0.607 for hollow/solid and homogeneity classification, respectively. The obtained hollow carbon spheres show a superior oxygen reduction electrocatalytic performance than solid spheres owing to their improved active site numbers. With the increasing structural complexity of nanomaterials, ML-driven synthetic design will endow great potential and promising prospects.