Expert-assisted and sub-unit-based molecular description enables ultra-fast and highly accurate prediction of the energy levels of non-fullerene acceptors in organic solar cells

Non-fullerene acceptors (NFAs) have recently emerged as an important class of materials enabling high-efficiency organic solar cells. However, most research activities on NFAs are carried out through a time-consuming trial-and-error experimental process with limited predictability. Therefore, there is a pressing need to develop a fast and efficient computational method to predict their properties, particularly their energy levels. Unfortunately, conventional computational or machine learning methods tend to produce large errors (0.2 - 0.5 eV) in energy level prediction. Moreover, the commonly used experimental method is also prone to large errors in determining the energy levels of NFAs. This paper introduces a novel ultra-fast machine learning method that accurately predicts NFA energy levels within only one second of computation on a common laptop computer. The method consists of two components: data cleansing and an expert-assisted sub-unit-based molecular description, which simplifies the complexity of molecular computation. Despite the limited data available for NFAs (less than a thousand), the method achieves a small average prediction error of 0.06 eV for energy levels, significantly outperforming known computational methods and even surpassing the classic experimental method, cyclic voltammetry (CV). This method is expected to accelerate NFA research and inspire the development of similar computation approaches to solve other material or molecular science problems.