Alfred Torsten Nordman: Parameters, Potentials and Probabilistic Methods in Molecular Modeling

Abstract

This thesis examines how physics-based molecular models can be constructed, calibrated, and evaluated under the practical limits imposed by computational cost, accessible simulation timescales, and the need for numerically stable and interpretable force fields, with particular focus on force fields as modular representations of molecular interactions. The work combines methodological foundations in computational chemistry, parameterization, molecular dynamics, sampling, Bayesian inference, and analysis with four studies that address complementary aspects of model performance. First, infrared spectra are used to assess force field quality. By comparing calculated and experimental infrared spectra for a large set of gas-phase compounds, the thesis shows that vibrational modes can serve as a useful test of how well molecular models reproduce measurable behavior. Second, a Bayesian framework is developed for three-point water models, in which force field parameters are inferred as probability distributions rather than point values. This enables uncertainty in the inferred parameters to be quantified and makes it possible to examine how this uncertainty relates to simulation error and limitations of the model class. Third, this framework is extended to polarizable water models, allowing the relation between increased physical detail, model behavior, and parameter uncertainty to be studied in a comparable setting. Finally, hydrogen halides are used as a test case for systematic comparison of force fields with different interaction forms and levels of complexity, in order to examine how individual modeling choices influence predictive performance. Taken together, the thesis contributes to a broader view of force field science in which force fields are treated not only as tools for simulation, but also as objects of systematic analysis. From this perspective, questions of functional form, parameterization, uncertainty, and intended application can be studied in a more explicit and structured way.