Interpretable machine learning for developing high-performance organic solar cells

Rapidly screening the underlying relationships between organic photovoltaics (OPVs) and their chemical structures remains an open challenge due to their complex interconnectivity. In this study, a new methodology for structure-property mappings of OPVs and device performances prediction is designed by combining the machine learning (ML) approach with the Taguchi Design of Experiments (TDOE). The established structure-property relationships are built up with the ML models from 240 data points of small molecule OPV systems and ten important microscopic features of OPVs. The quite remarkable performance of the ML model (Pearson's coefficient = 0.79) depicts its ability to extract hidden physical principles of OPVs. The TDOE model shows that molecular orbitals other than the highest and the lowest ones that are not frequently considered in the designing process of OPVs play quite essential roles in developing promising OPV materials. Moreover, strategies to boost the design of high-performing devices with different values of the considered features are also extracted from the model with the DOE approach. These results reveal that ML combined with DOE is an impressive package for guiding the design process effectively and efficiently. • New design guidelines for small molecule organic solar cell materials developed using interpretable machine learning. • The design model can predict the underlying physical phenomena between considered features. • The relations between orbitals rather than just HOMO and LUMO and optical bandgap were investigated.