Physics-Aware Learning of Donor–Acceptor Pair Interactions for High-Efficiency Organic Photovoltaics

Predicting functional performance in complex organic and polymer-based electronic materials remains a key challenge for data-driven materials design, as device-level properties are governed by coupled molecular interactions rather than isolated components. In organic photovoltaic (OPV) systems, this challenge is particularly pronounced when targeting high power conversion efficiency (PCE) regimes. In this work, we present a physics-aware learning framework modeling donor–acceptor pair interactions using a combination of physically informed molecular descriptors and ensemble learning strategies. By integrating global electronic and structural descriptors with pairwise interaction features, the proposed model achieves improved predictive fidelity across multiple photovoltaic properties and demonstrates enhanced reliability for high-PCE systems. To address the intrinsic data imbalance in OPV data sets, a screening-oriented training strategy incorporating targeted oversampling of high-performance samples is further employed, leading to a substantial improvement in prediction accuracy in the high-PCE region. High-throughput virtual screening based on the trained model identifies a set of promising donor–acceptor combinations with high predicted PCEs, providing practical guidance for experimental synthesis and device optimization. More broadly, this study demonstrates how interaction-aware representations and task-oriented learning strategies can enhance the reliability of data-driven models for complex materials discovery.