Decoding Phases of Matter by Machine-Learning Raman Spectroscopy

Phase transitions of condensed matter have long been a spotlight issue studied by extensive theoretical and experimental investigations. Machine learning can build an integral model-dominant workflow to statistically analyze the collective dynamics of materials and deduce the structure. We use a supportvector-machine algorithm to propose an effective method to recognize the orthorhombic, tetragonal, and cubic phases as well as to construct the phase diagram in ferroelectric crystals by mining and learning the behavioral vectors of the phonon vibrations in a crystalline lattice from Raman scattering, which is a tool typically used to detect structural properties at the molecular level. This study creates a unifying framework including material synthesis and characterization, feature engineering and principal-component analysis, learner evaluation and optimization, structure prediction, and future development of the model. It paves the way to the application of a generic approach for predicting unexplored structures and materials in the future.