Dynamical Heterogeneity in Supercooled Water and its Spectroscopic Fingerprints

TL;DR

This work uses a SCAN-DFT machine-learning interatomic potential to study dynamical and spectroscopic signatures of LDL and HDL water in the deeply supercooled regime. It reveals that LDL is slower and dynamically heterogeneous, with long-lived caging and a population of dormant molecules, while HDL remains more diffusive and uniform. Infrared spectra show LDL-specific features, notably a blue-shifted, sharper libration band and enhanced cross-correlation contributions, signaling stronger collective hydrogen-bond dynamics. Together, these findings provide microscopic dynamical and spectroscopic fingerprints to guide experimental detection of the LLCP in supercooled water and link structural fluctuations to dynamical heterogeneity.

Abstract

A growing body of theoretical and experimental evidence strongly supports the existence of a second liquid-liquid critical point (LLCP) in deeply supercooled water leading to the co-existence of two phases: a high-and low-density liquid (HDL and LDL). While the thermodynamics associated with this putative LLCP has been well characterised through numerical simulations, the dynamical properties of these two phases close to the critical point remain much less understood. In this work, we investigate their dynamical and spectroscopic features using machine-learning interatomic potentials (MLIPs). Dynamical analyses using the van-Hove correlation function, reveal that LDL exhibits very sluggish and heterogeneous molecular mobility, in contrast to the faster and more homogeneous dynamics of HDL. Infrared absorption (IR) spectra further show clear vibrational distinctions between LDL and HDL, in particular in the far IR region between 400 - 1000 cm-1. Together, these findings provide new dynamical fingerprints that clarify the microscopic behavior of supercooled water and offer valuable guidance for experimental efforts aimed at detecting the long-sought liquid-liquid transition.

Figures (6)

• Figure 1: Mean squared displacement as a function of time for LDL (blue) and HDL (orange) water. While both phases exhibit diffusive behavior at long times, the HDL phase shows significantly faster molecular mobility.
• Figure 2: Self part of the van Hove correlation function $G_S(r,t)$ for ambient liquid water (left), HDL (middle) and HDL (right) at selected time intervals. Liquid water and HDL phase exhibits a smooth shift in the peak over time. In contrast, LDL shows a slower and more complex evolution, with pronounced tail and different peaks at intermediate timescales, indicative of dynamic heterogeneity.
• Figure 3: (a) Displacement time series $d(t)$ for a dormant (blue) and an active (orange) water molecule. The dashed line marks the jump threshold at $2.5 \, \text{\AA}$. Peaks above this threshold correspond to jump-like events. (b) Snapshot of the simulation box. Molecules classified as dormant are shown in blue; active molecules are shown in orange. Roughly 30 % of the molecules are dynamically arrested, indicating spatial heterogeneity in mobility.
• Figure 4: Histogram of the time-averaged number of hydrogen bonds per molecule. The total distribution is decomposed into contributions from dynamically dormant (blue) and active (orange) molecules. Active molecules tend to be undercoordinated and exhibit a broader distribution with a tail toward fewer hydrogen bonds.
• Figure 5: Infrared absorption spectra of ambient, LDL, and HDL water compared with experimental data. Simulated IR spectra computed using dipole moments at the SCAN-DFT level for ambient liquid water, LDL, and HDL. The spectra reveal distinct features across the frequency range, with particularly marked differences in far-IR region. The comparison with experimental data Bertie1996 validates the accuracy of the SCAN functional and the MLIP in capturing spectroscopic features of liquid water.