Decoding Phases of Matter by Machine-Learning Raman Spectroscopy

Exploring the phases of matter is important in condensed matter physics and materials discovery, and here Raman spectroscopy gives a material's structural fingerprint. The authors develop a highly correlated kernel model based on a set of Raman-scattering spectra, to deduce thermally induced transitions among three phases in a model ferroelectric crystal, by mining and learning the phonon behaviors in the crystal lattice. The underlying physical mechanism behind these behaviors, and just how the model ``learns'' these rules to make precise predictions, are clarified. This insight paves the way to applying the generic approach for prediction of unexplored structures and materials.