🔬 condensed matter

Dynamical Heterogeneity in Supercooled Water and its Spectroscopic Fingerprints

Using machine-learning interatomic potentials, this study reveals that the low-density liquid phase of supercooled water exhibits sluggish, heterogeneous dynamics and distinct far-infrared vibrational signatures compared to the high-density phase, providing new spectroscopic fingerprints to guide experimental detection of the liquid-liquid transition.

Original authors: Cesare Malosso, Edward Danquah Donkor, Stefano Baroni, Ali Hassanali

Published 2026-04-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Cesare Malosso, Edward Danquah Donkor, Stefano Baroni, Ali Hassanali

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine water not just as the clear liquid you drink, but as a shape-shifting character with a secret double life. We all know water freezes into ice, but what happens when you cool it down below freezing without letting it turn into solid ice? Scientists call this "supercooled" water, and for decades, they've suspected it has a hidden personality crisis: it can exist as two different types of liquid at the same time.

This paper is like a detective story where the authors use super-smart computer simulations to catch these two "liquid twins" in the act and figure out exactly how they behave differently.

Here is the breakdown of their discovery in plain English:

The Two Liquid Twins: HDL and LDL

The researchers believe that deep in the supercooled zone, water splits into two distinct phases:

HDL (High-Density Liquid): Think of this as the busy, chaotic city. The molecules are packed tight, but they are moving around quickly, bumping into each other, and getting things done. It's like a crowded dance floor where everyone is dancing energetically.
LDL (Low-Density Liquid): This is the quiet, frozen-in-time village. The molecules are more spread out, but they are incredibly sluggish. They are stuck in place, barely moving, like people frozen in a statue pose.

The "Cage" Analogy: Why LDL is So Slow

To understand the difference, imagine a molecule is a person in a room.

In HDL (The City): The person is in a crowded room, but the crowd is moving. If you want to walk across the room, you might bump into someone, but you can wiggle through and keep moving. You are "diffusing" (moving freely).
In LDL (The Village): The person is trapped in a cage made of their neighbors. They can wiggle their arms and shake a little, but they can't actually walk anywhere. They are "dynamically arrested."

The paper found that in the LDL phase, about 30% of the water molecules are completely stuck for hundreds of nanoseconds (which is an eternity in the world of atoms). They are "dormant." Meanwhile, the other 70% are "active" and suddenly burst into motion, jumping from one cage to another. This creates a patchwork of movement: some parts of the liquid are frozen, while others are jumping around. This is called Dynamical Heterogeneity.

The "Fingerprint" Test: Listening to the Water

How do we know these two liquids are different if we can't see them? The authors used a technique called Infrared Spectroscopy. Think of this as giving the water a "voice test." Every time a molecule vibrates or wiggles, it sings a specific note.

The High Notes (OH Stretching): When they looked at the high-pitched sounds (vibrations of the hydrogen bonds), the two liquids sounded very similar. It was hard to tell them apart just by listening to the high notes.
The Low Notes (The Libration Band): This is where the magic happened. When they listened to the low, rumbling sounds (which represent how molecules try to rotate or spin but get blocked by their neighbors), the difference was huge.
- HDL had a broad, fuzzy low note, like a drum being hit loosely. It meant the molecules had freedom to spin.
- LDL had a sharp, high-pitched, clear low note. This meant the molecules were in a stiff, rigid cage. They were trying to spin, but the "walls" of their hydrogen-bond cage were so strong and organized that they couldn't move much.

The Big Reveal

The most important finding is that the difference between these two liquids isn't just about how they are arranged (structure); it's about how they move together (dynamics).

The LDL phase is special because its molecules are locked in a collective, rigid dance. They aren't just stuck individually; they are stuck together in a way that creates a "ferroelectric" feel (like a tiny magnet where everything points the same way). The paper suggests that if you want to find the "Liquid-Liquid Transition" in a real lab experiment, you shouldn't just look at the structure; you should listen to those low-frequency rumbling sounds. That is the secret fingerprint that tells you which "liquid twin" you are looking at.

Why Does This Matter?

Water is weird. It expands when it freezes, it gets denser when it cools, and it behaves strangely under pressure. This paper helps explain why. It suggests that water's weirdness comes from this tug-of-war between the "busy city" (HDL) and the "frozen village" (LDL).

By using advanced AI (Machine Learning) to simulate water with extreme accuracy, the authors have provided a roadmap for experimentalists. They are essentially saying: "If you want to prove that water has two liquid forms, stop looking at the high notes and start listening to the low, rumbling vibrations. That's where the truth is hiding."

1. Problem Statement

Supercooled water exhibits anomalous thermodynamic and dynamic properties that are hypothesized to stem from a Liquid-Liquid Critical Point (LLCP), leading to the coexistence of two distinct liquid phases: High-Density Liquid (HDL) and Low-Density Liquid (LDL). While the thermodynamic existence of this LLCP has been supported by numerical simulations, the dynamical properties of these two phases near the critical point remain poorly understood.

Key gaps in current knowledge include:

Lack of Experimental Confirmation: Direct observation of the LLCP is hindered by rapid crystallization ("no-man's land"), and indirect experimental evidence (e.g., IR spectroscopy) lacks systematic theoretical support for interpretation.
Limitations of Classical Models: Traditional empirical force fields are inadequate for accurately describing high-frequency vibrational spectroscopy and subtle quantum mechanical effects required to distinguish HDL and LDL dynamics.
Microscopic Mechanism: The specific nature of dynamical heterogeneity (spatial and temporal variations in mobility) and its link to structural defects (hydrogen bonding) in supercooled water requires a more rigorous, quantum-accurate description.

2. Methodology

The authors employed Machine Learning Interatomic Potentials (MLIPs) to achieve quantum mechanical accuracy at a computational cost comparable to empirical potentials.

Potential Model: They utilized a Deep Potential (DeePMD) model trained on high-quality Density Functional Theory (DFT) data using the SCAN (Strongly Constrained and Appropriately Normed) meta-GGA functional.
Simulation Setup:
- System: 1,536 water molecules.
- Conditions: NPT simulations were run to equilibrate HDL (235 K, 3600 bar) and LDL (235 K, 2800 bar). Subsequent NVT simulations were performed at 235 K for both phases (100 ns for HDL, 400 ns for LDL) and at 330 K for ambient water.
- Software: LAMMPS with DeePMD-kit.
Dynamical Analysis:
- Mean Squared Displacement (MSD): To calculate diffusion coefficients.
- Van Hove Correlation Function ( $G_S(r,t)$ ): To analyze single-particle displacement distributions and detect non-Gaussian behavior indicative of dynamical heterogeneity.
- Non-Gaussianity Parameter ( $\alpha_2(t)$ ): To quantify deviations from Gaussian diffusion.
- Jump Detection: A protocol involving signal filtering (Butterworth filter) and smoothing was applied to identify "jump-like" translational motions. Molecules were classified as "dormant" (no jumps > 2.5 Å over 400 ns) or "active."
- Hydrogen Bond Analysis: Time-averaged H-bond counts were computed using geometric criteria to correlate structure with dynamics.
Spectroscopic Analysis:
- Infrared (IR) Spectra: Computed as the Fourier transform of the total dipole moment time correlation function.
- Decomposition: Spectra were decomposed into self (intra-molecular) and cross (inter-molecular) correlation terms to isolate collective effects.
- Dipole Calculation: Molecular dipoles were predicted by a deep neural network trained on SCAN-DFT reference data using Maximally Localized Wannier Functions (MLWFs).

3. Key Results

A. Dynamical Heterogeneity and Mobility

Diffusion Coefficients: HDL exhibits significantly faster diffusion ( $D \approx 2.77 \times 10^{-3}$ Å $^2$ /ps) compared to LDL ( $D \approx 1.7 \times 10^{-5}$ Å $^2$ /ps). LDL mobility is suppressed by two orders of magnitude.
Van Hove Analysis:
- HDL: Shows homogeneous, diffusive dynamics with a single peak shifting smoothly over time.
- LDL: Exhibits pronounced dynamical heterogeneity. The Van Hove function develops "fat tails" and shoulders, indicating a coexistence of mobile and arrested particles. The peak position remains nearly static even after 37 ns, while a tail of mobile particles extends far.
Non-Gaussianity: The non-Gaussianity parameter $\alpha_2(t)$ decays rapidly for ambient water and HDL but remains persistently high ( $\approx 4$ ) for LDL even at 40 ns, confirming non-Gaussian, heterogeneous dynamics.
Active vs. Dormant Molecules:
- Approximately 30% of LDL molecules are "dormant" (dynamically arrested) over the 400 ns simulation, showing no translational jumps.
- Dormant molecules show no significant orientational changes, while "active" molecules exhibit coupled translational jumps and rotational rearrangements.
- Dormant molecules show a slight tendency toward higher H-bond counts, but the distinction is not absolute, suggesting mobility arises from a complex interplay of local environment and collective rearrangements rather than static defects alone.

B. Spectroscopic Fingerprints (IR Spectra)

High-Frequency Region (OH Stretch, ~3400 cm $^{-1}$ ): LDL shows a red-shift of ~40 cm $^{-1}$ compared to HDL, indicating stronger hydrogen bonds. However, the spectra largely overlap, making this region less distinct for phase identification.
Low-Frequency Region (Libration, 400–1000 cm $^{-1}$ ):
- LDL: Exhibits a sharper, blue-shifted libration peak (shifted by ~140 cm $^{-1}$ relative to HDL) with a narrower Full Width at Half Maximum (FWHM).
- HDL: Shows a broader, red-shifted response, indicating greater rotational freedom and a disordered environment.
Intermolecular Stretching (~200 cm $^{-1}$ ): LDL shows a more intense and slightly blue-shifted peak, reflecting stiffer hydrogen-bond cages.
Self vs. Cross Correlation: The spectral differences in the low-frequency region are predominantly driven by cross-correlation terms (intermolecular dipole-dipole interactions). This confirms that the distinct vibrational response of LDL is a result of collective, large-scale spatial correlations rather than just local molecular structure.

4. Significance and Impact

New Dynamical Fingerprints: The study provides the first theoretical IR spectra distinguishing HDL and LDL, identifying the libration band (400–1000 cm $^{-1}$ ) as the critical region for experimental detection of the liquid-liquid transition.
Mechanistic Insight: It clarifies that the sluggish dynamics in LDL are not merely due to static structural defects but are governed by collective hydrodynamic effects and long-lived polarization fluctuations, with a significant fraction of the liquid becoming dynamically arrested.
Methodological Advancement: The work demonstrates the capability of MLIPs trained on high-level DFT data to resolve subtle dynamical and spectroscopic features that classical force fields miss, bridging the gap between quantum accuracy and the long timescales required to observe supercooled water phenomena.
Experimental Guidance: The findings offer specific targets for future experiments (e.g., Correlated Vibrational Spectroscopy) to detect the elusive liquid-liquid transition by focusing on low-frequency collective modes and dynamical heterogeneity.

In conclusion, this paper establishes that the LDL phase of supercooled water is characterized by a unique state of dynamical arrest and collective rigidity, which manifests as distinct, measurable signatures in the low-frequency infrared spectrum.