Prediction of Nanoparticle Dispersion by Machine Learning Using Various Molecular Descriptors

Surface-modified nanoparticles (NPs) have attracted major interest in widespread applications due to their good dispersibility in solvents. Although the selection of modifiers and solvents is also crucial to achieving good dispersibility, experimental screening consumes a significant amount of financial, material, and time resources due to the many candidates of modifiers and solvents. In this work, supervised machine-learning models were established to screen modifiers and solvents while using Hansen solubility parameters (HSPs), conductor-like screening model (COSMO)-based molecular information, and molecular access system keys as explanatory variables. Binary classification problems (dispersion or not) were solved while aligning the number of dispersion data by each modifier. As a result, this oversampling procedure allowed a stable and accurate prediction of dispersion for hydrophilic and hydrophobic NPs. Additionally, HSPs allowed the highest accuracy of 84.3%, recall of 85.9%, and F1-score of 84.0%, in which the feature importance analysis supported HSPs as useful molecular descriptors. These results suggest that a machine-learning approach using HSPs can be an appealing tool to predict the dispersibility of surface-modified NPs and to computationally screen the modifier and solvent.