Prediction of global Raman spectral features in minerals from structure and composition

How well structure and composition predict the global features of mineral Raman spectra remains unclear. Using 1838 quality-controlled RRUFF spectra, 11 spectral endpoints and five descriptor classes, we evaluated formula-grouped predictive models. The full model achieved out-of-fold R² of 0.38-0.44 for log peak count, spectral centroid and skewness, and 0.25 for log median linewidth. Relative to a chemical-family and crystal-system baseline, the descriptors improved prediction for 8 of 11 endpoints, most strongly for median linewidth (ΔR² = +0.23). Group-theoretical Raman-active mode count was only weakly associated with observed peak count (Spearman ρ = 0.33), and leave-one-family-out validation showed substantial losses in transfer performance. These results quantify both the predictive content and the extrapolation limits of static structure- and composition-derived descriptors.