Interpretable machine learning reveals morphology-mediated performance pathways in organic solar cells

Abstract

Organic solar cell (OSC) performance arises from coupled material, electronic, processing, and morphology-dependent mechanisms, yet the transferability of these relationships across heterogeneous literature-derived datasets remains uncertain. Here, interpretable machine learning (ML) was applied to a curated dataset of 476 unique OSC devices from 73 publications to predict Voc, Jsc, and fill factor (FF) and to evaluate the relative contribution of structural/process and electronic descriptors. Random, DOI-grouped, and donor–acceptor-pair-grouped cross-validation were used to distinguish within-literature interpolation from cross-publication and material-combination transfer. FF exhibited the strongest transferable predictive signal, retaining R2=0.450 under DOI-grouped validation and R2=0.568 for unseen donor–acceptor combinations, whereas Voc and Jsc showed weaker generalization. Active-layer thickness emerged as the dominant morphology-related predictor, particularly for FF, consistent with its influence on transport distance, recombination, and charge extraction. Electronic descriptors provided target-dependent incremental information but did not consistently improve cross-publication performance. These results demonstrate that applicability-aware, physics-interpretable ML can distinguish interpolation from transferable structure–processing–performance relationships and provide a more reliable basis for OSC device optimization.