Data-driven spectrum interpretation in the liquid phase

Abstract

Interpreting liquid-phase electronic spectra is challenging due to the complex interplay of atomic motion and intermolecular interactions. These effects create a wide distribution of local atomistic environments, each of which can yield a signifcantly different spectrum. Because experiments only capture ensemble averages, any observed spectral change refects a latent shift in the underlying distribution of spectrally relevant structural features. In this thesis, I introduce a data-driven framework for disentangling the structure–spectrum relationship by adapting the recently proposed emulator-based component analysis. Applied to extensive simulated datasets, this spectrum-variance-driven dimensionality reduction method transforms the high-dimensional problem into a low-dimensional form, uncovering structural trends behind spectral changes. Across the studied cases, most of the spectral variation is explained by a small fraction of structural variance, represented by only a few latent structural degrees of freedom. This emerging pattern may indicate a general principle in spectroscopy of liquids.