Through Rainbow-Tinted Glasses: Machine Learning-Driven Modeling of Chromophores

Doctoral thesis, 2026

Chromophores are a class of molecules that give color to the world around us, from the chlorophyll in plants that enables photosynthesis to the retinal molecules in our eyes that allow us to see. Chromophores are also fundamental for developing a range of technologies crucial for a transition to a sustainable society, including photovoltaics and energy storage. While chromophores have been widely studied experimentally, we still lack a sufficient understanding of their structure and dynamics on the atomic scale. In particular, many chromophores are glass formers or form supramolecular aggregates due to intermolecular interactions, leading to a complicated landscape of interactions spanning many time and length scales. Addressing this limitation requires atomistic simulations capable of connecting microscopic dynamics to experimentally measurable quantities. In this spirit, this thesis presents a simulation framework that links electronic structure calculations via molecular dynamics simulations to experiments, with a focus on neutron scattering.

The key ingredient in this framework is machine-learned interatomic potentials, enabling simulations with the accuracy of quantum mechanical calculations for large systems of chromophores. Methodological developments focus on the neuroevolution potential framework implemented in the GPUMD package. A major contribution is the development of the calorine package, which is a companion software for GPUMD that interfaces with the broader scientific software ecosystem.

The framework is applied to three challenging applications: glass formation, optical response, and neutron scattering. First, glass formation occurs beyond timescales accessible to molecular dynamics simulations, and this limitation is circumvented by extrapolating relaxation processes from nanoseconds to experimental timescales using Bayesian regression. Second, optical properties are obtained by using the neuroevolution potential framework to predict tensorial properties such as the dipole moment, or spectral quantities such as the electronic dielectric function. Finally, instrument-specific inelastic neutron scattering signatures are predicted using electronic structure calculations, machine learning, and correlation functions. Together, these developments establish a framework for connecting atomistic simulations with experimental observables, enabling modeling of chromophores over multiple time and length scales. The framework is transferable and directly applicable to other systems.