Machine learning for spectroscopy (ML4SPEC)

Spectroscopy is an indispensable technique in science and engineering for studying the properties of molecules and materials, essential for the complete characterization of complex multi-component and multi-dimensional materials. By varying the source energy in these experiments, researchers can access the internal states and dynamics of physical systems, allowing for in-depth analysis of the material. For example, valence and conduction band spectra are crucial for solar cells, spin spectra for quantum computing devices, optical spectra for organic electronics, vibrational spectra for catalysis, conductivity spectra for lightweight batteries, X-ray spectra for novel semiconductors, and infrared and nuclear magnetic resonance spectra for pharmaceutical characterization. Quantum mechanical calculations are required for the quantitative interpretation, prediction, and understanding of material properties based on complex spectral features. However, state-of-the-art calculations are limited due to incomplete treatment of several critical factors: 1) correlations between electrons, 2) temperature effects, 3) non-pristine crystals with structural or chemical disorder, 4) initial and final states of the probe particles, and 5) external applied fields. Full quantitative treatment of these properties depends on accurate calculations, which remain computationally expensive, often requiring hundreds to thousands of CPU cores on high-performance computing (HPC) clusters for materials or molecules of even moderate structural complexity. To accelerate the computation of material properties, Machine Learning (ML) is increasingly employed to compute structural and ground state properties and is gradually gaining traction for excited state calculations and quantitative spectroscopy. ML has the potential to revolutionize quantitative spectroscopy, moving away from manual analysis and interpretation towards automated, algorithm and data-driven methodologies.