Learning the limits: how data, diversity, and representation control machine-learning predictions of reorganisation energy

Abstract

Accurate and scalable prediction of hole and electron reorganisation energies (λ_h and λ_e) is a persistent bottleneck in the data-driven design of organic semiconductors, as routine ab initio calculations remain impractical for large molecular libraries. This work presents a systematic and interpretable evaluation of how molecular representation, chemical diversity, and dataset size constrain the accuracy and transferability of machine-learning models for predicting λ_h and λ_e. Three complementary datasets are analysed: a chemically diverse benchmark of approximately 5000 molecules with paired λ_h and λ_e values, a thiophene-focused dataset comprising 1486 molecules, and a sequence of progressively augmented datasets extending to nearly 13 000 structures. Fifteen molecular descriptor schemes and twelve learning algorithms, spanning linear, kernel-based, ensemble, and graph-based models, are benchmarked under consistent training and validation protocols. Across broad chemical space, predictive performance is primarily governed by molecular representation, with hybrid descriptors that combine RDKit features and multiple molecular fingerprints consistently outperforming single-source encodings, while graph neural networks underperform in highly diverse regimes. Constraining chemical diversity leads to substantial accuracy gains, particularly for electron reorganisation energies, whereas increasing dataset size improves robustness and generalisation with rapidly diminishing returns beyond modest augmentation. Model interpretation using SHAP analysis reveals stable and physically meaningful design trends across all datasets, showing that rigid, extended π-conjugation, low conformational flexibility, and balanced charge distribution systematically reduce reorganisation energies. These results define realistic performance limits for machine-learning prediction of reorganisation energy and provide concrete guidance on representation choice, dataset design, and molecular optimisation strategies for high-mobility organic electronic materials.