From SEM morphology to PEC behavior: machine-learning identification of a UV-Vis spectral descriptor for BiVO₄...

In this study, we demonstrate that the nanoscale morphology of BiVO₄ photoanodes can be quantitatively translated into optical absorption features, and further into PEC behavior, through a combined convolutional neural network (CNN) and Gaussian Mixture Regression (GMR) framework. CNN models successfully predicted UV-Vis spectra directly from SEM images, showing that multiscale morphological information-especially subsurface contrast captured by BEI-is correlated with two major interband transitions (Bi 6s → V 3d, O 2p → V 3d). Using these spectra as descriptors, the GMR model reconstructed full J-V curves with high accuracy, enabling PEC performance prediction from optical data alone. Scrutinizing the relation between the J-V and UV-Vis curves further identified a single spectral descriptor: the inflection point near ~ 360 nm, which consistently distinguished high- and low-performing samples in the spectra and corresponds to the O 2p → V 3d transition. This trend was also confirmed in the experimentally measured spectra of newly prepared samples. This feature is tentatively attributed to oxygen-related electronic structure perturbations associated with morphology-dependent variations in local coordination and disorder. Although the descriptor was identified for BiVO₄ photoanodes prepared by a single MOD route, this work shows that UV-Vis spectra, despite their simplicity, contain sufficient encoded morphological information to act as a fast, nondestructive probe for PEC measurements.