Machine learning with quantum chemistry descriptors: predicting the solubility of small-molecule optoelectronic materials for organic solar cells

Abstract

Solubility prediction is important in developing high-performance optoelectronic materials for organic solar cells, and can assist the synthetic route and chemical process design of optoelectronic materials, and control the morphology of bulk-heterojunctions. Here we report a successful approach that can effectively predict the solubility of optoelectronic materials in any solvents by using a combination of machine learning and quantum chemistry descriptors. Temperature combined with quantum chemistry calculated molecular vdW surface area (area), positive electrostatic potential (ESP) variance (σ₊²) and negative ESP variance (σ₋²) were used as a small set of descriptors containing only 7 bits of data. It is the smallest set of descriptors currently used and shows good predictive performance to predict the solubility. This small set of descriptors enables us to predict the solubility of any small molecule in various solvents with a small number of quantum chemical calculations. The solubility of PCBM and Y6 in 42 common solvents used in organic chemistry was predicted, and 10 solvents with the highest solubility are screened out from the dataset. This model can be applied to other small-molecule systems to rapidly predict their solubility in any solvent and provide an important parameter for designing promising high-performance optoelectronic materials for organic solar cells.