Interpretable and Leakage‐Controlled Fingerprint Prediction of Organic Solar‐Cell Donor–Acceptor Pair Efficiency With Structure‐Resolved Interpretation and Fingerprint‐Only Data Integration

Predicting power conversion efficiency (PCE) of donor–acceptor (D–A) pairs in organic solar cells remains challenging due to heterogeneous labels, limited structural coverage, and representation leakage. We develop a fingerprint‐based workflow that separates structure‐resolved interpretation from prediction‐oriented integration of external fingerprint‐only data. From 836 in‐house D–A pairs, scaffold‐disjoint diagnostic and quarantine sets were excluded, leaving 715 pairs for target‐blind, group‐aware modeling. A Random Forest baseline achieved Pearson on the fixed test set but showed limited scaffold extrapolation (). After radius‐2 fingerprint alignment with 535 Gao records, a 1,250‐sample role‐aware TextCNN ensemble improved prediction to and . Nested resampling yielded robust performance (, ), with partition variability dominating initialization variability. SHAP interpretation was restricted to the structure‐resolved branch. The results demonstrate the value of fingerprint‐only data integration while highlighting limited scaffold‐level generalization.