Abstract
Water is an exceptionally adaptable liquid whose behaviour across temperature, pressure, composition, and interfaces underlies processes ranging from planetary interiors to electrochemical energy storage and atmospheric chemistry. Many of its most important properties—dielectric screening, vibrational spectra, proton transport, and interfacial molecular orientation—are emergent, collective responses rather than singlemolecule attributes, and they cannot be inferred from local structure alone.
This thesis couples ab initio molecular dynamics with machine-learning models of the potential energy, the molecular dipole, and the polarizability to connect electronic structure to structural, dielectric, transport, and spectroscopic observables. Nuclei are treated classically throughout.
Four questions are addressed. First, at 1000 K and densities of 0.88–1.57 g cm−3 (roughly 1–10 GPa), four exchange–correlation approximations—PBE, PBE-D3, RPBED3, and SCAN—converge more closely for pair structure, self-diffusion, and vibrational spectra than at ambient conditions, yet proton-transfer event counts remain model-dependent. Agreement in equilibrium structure therefore does not guarantee ii transferable reactivity.
Second, the dielectric response is examined in two regimes. For molecular water, six trajectory models combined with three dipole models yield fluctuation permittivities consistently in excess of the experimental total, and the spread across trajectories exceeds that across dipole models. For dissociated water at 2000 K and 10–30 GPa, a Wannier total-current formulation yields conductivities that increase monotonically with pressure while the dielectric increment correspondingly declines.
Third, in KOH and NaOH from 1 to 15 M, hydroxide hydration shifts from OH−(H2O)4 toward OH−(H2O)3, while the KOH proton-transfer barrier increases only marginally. A sampled sequence suggests a triggered hopping pathway, and first-principles Raman spectra reproduce the concentration dependence, with the NaOH hydroxide band appearing at higher frequency than that of KOH.
Fourth, machine-learning models extend sum-frequency generation calculations to a 16,384-molecule water–air interface over nearly 1 ns, resolving how the response builds up with depth. Intermolecular cross-correlations couple the outermost and subsurface layers, so the low-frequency positive O–D shoulder is not confined to the top layer.
These results demonstrate that electronic accuracy, sampling, and observable definitions must be validated jointly, and that collective aqueous observables cannot be reconstructed from a single local descriptor.