Interpretable Prediction of Full UV-Vis Absorption Spectra of Organic Molecules in Solution

Accurate prediction of experimental UV–Vis absorption spectra is essential for understanding molecular electronic structures and guiding the design of organic optical materials. However, predicting full absorption spectra, particularly those containing multiple absorption bands and solvent-dependent spectral features, remains challenging. Here, we develop a hierarchical molecular graph attention network (HM-GAT) for predicting full UV–Vis absorption spectra of organic molecules in solution directly from two-dimensional molecular structures. HM-GAT combines hierarchical molecular representations with substructure–solvent attention to capture solvent-dependent absorption behavior and provide chemically meaningful interpretation. A threespectral-window architecture defined on the photon-energy scale and a transfer-learning strategy are employed to improve predictive accuracy. By pretraining on 84,383 molecule–solvent pairs with experimental optical properties and fine-tuning on 12,788 experimental spectra, HM-GAT achieves a test-set spectral mean absolute error of 0.060 over 330–1000 nm, accurately reproducing multiple absorption bands, molar extinction coefficients, and solvent-dependent spectral shifts. The three-spectral-window architecture reduces prediction errors for multi-peak spectra by 25.0% compared with the single-window configuration. Furthermore, substructure attention scores exhibit qualitative agreement with natural transition orbital distributions, providing insights into chromophoric substructures associated with electronic absorption. Independent evaluation using representative organic optical materials excluded from both modeldevelopment datasets further supports the generalization capability of HM-GAT. These results establish a scalable and chemically interpretable framework for virtual molecular spectroscopy and computational screening of organic molecules with tailored optical absorption properties.