Interpretable machine-learning for predicting power conversion efficiency of non-halogenated green solvent-processed organic solar cells based on Hansen solubility parameters and molecular weights of polymers

Abstract

The green-solvent-processable strategies (e.g., open-air-printable ink for environmentally-friendly fabrication) are crucial to enable sustainable manufacturing of large-scale devices with safety considerations (e.g., uncomplicated handling, low hazard potential, and high availability) to the commercialization of Bulk heterojunction (BHJ) based organic solar cells (OSCs). Unfortunately, identifying the relationship between the performances of non-halogenated green solvent-processed BHJ-based OSCs and arbitrary polymer-solvent pairs has been a challenging problem for the profound sustainable development of the OSCs field because of the lack of reliable quantitative and accurate prediction methods. To address this problem, we show that the Gradient Boosting (GB) machine-learning model, trained on several types of potentially relevant descriptors (e.g., polymer molecular weight (MW) and Hansen solubility parameters (δd, δp, δh)), can be used to predict power conversion efficiency (PCE) of non-halogenated green solvent-processed OSCs, thereby providing a novel data-driven paradigm to select appropriate polymer-solvent pairs. The feature importance analysis, generated by the SHAP method, reveals that polymer MW has a profound impact on the prediction of PCE of non-halogenated green solvent-processed BHJ-based OSCs. The GB machine-learning model constructed here is capable of extracting the highly non-linear polymer/non-halogenated green solvent pairs-performance mapping of BHJ-based OSCs, and the SHAP feature importance analysis can be easily applied to strengthening the relationship between the complex machine-learning models and the scientists in the experimental device optimization of non-halogenated green solvent-processed BHJ-based OSCs.