Machine Learning-Driven Accelerated Discovery and Performance Prediction of High-Efficiency Non-Fullerene Acceptors for Organic Solar Cells

Abstract

Currently, the rapid development of organic solar cells is constrained by the vast chemical space of potential active layer materials, leading to increasingly low efficiency in traditional trial-and-error experimental synthesis. This study proposes a comprehensive machine learning framework aimed at accelerating the discovery of high-performance non-fullerene acceptors. By combining high-throughput density functional theory calculations with a curated dataset derived from experimental literature, we develop a multi-stage pipeline incorporating molecular fingerprinting and feature engineering. Through evaluation of various machine learning algorithms, XGBoost, and graph neural networks, we predict key photovoltaic parameters (e.g., power conversion efficiency and highest occupied molecular orbitals), identify key molecular fragments and electronic descriptors controlling charge dissociation and open-circuit voltage. This interpretability provides actionable design rules for molecular engineering. This research demonstrates that combining machine learning with computational chemistry can significantly reduce the time and economic cost of developing organic solar cells, providing a robust paradigm for data-driven design of next-generation optoelectronic materials.