Glassy Arrest Behind the Apparent Second Liquid in Water

Florian Pabst, Ali Hassanali · 2026

Read Glassy Arrest Behind the Apparent Second Liquid in Water by Florian Pabst, Ali Hassanali (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-00794.txt.

Introduction

Pabst and Hassanali challenge the usual reading of water's strange behavior as proof of two distinct liquids. Using machine-learning-accelerated simulations checked against spectroscopic, structural, and dynamical experiments, they show that supposed two-liquid signals match the onset of a glassy, non-ergodic state instead. Their mapping across pressure and temperature points to a simpler reading. Low-density water may simply have a glass-transition temperature of 189 K, plus or minus 8.

Catalog record

CREATORFlorian Pabst, Ali Hassanali
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49434
LICENSEhttp://creativecommons.org/licenses/by-sa/4.0/
RIGHTSCC BY SA
SOURCEarXiv

Source: https://arxiv.org/abs/2604.00794

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Glassy Arrest Behind the Apparent Second Liquid in Water Florian Pabst, Ali Hassanali Abstract The origin of water’s anomalous behavior remains a central open problem in the physical sciences and is often attributed to a liquid–liquid transition (LLT) between high- and low-density liquid states deep in the supercooled regime. Experimental access to this region has been challenging due to rapid crystallization, leaving atomistic simulations as a major source of supporting evidence. Using extensive machine-learning–accelerated first-principles simulations in direct comparison with spectroscopic, structural, and dynamical experimental measurements, we show that features commonly interpreted as signatures of two-liquid behavior emerge at the onset of dynamical arrest. Specifically, we find that two-state fluctuations previously associated with an LLT reflect a transformation from a high-density liquid to a kinetically arrested low-density glass. By mapping equilibrium dynamics across pressure and temperature, our results suggest a reassessment of water’s metastable landscape, in which apparent two-state behavior may reflect a relatively high glass-transition temperature of ambient-pressure low-density water, 189 ± \pm 8 K—remarkably close to the temperature previously associated with the LLT. Water defies the expectations of a simple liquid. Upon cooling, it expands rather than contracts, while its compressibility and heat capacity exhibit pronounced extrema. These anomalies have long hinted at an underlying organizing principle that remains unresolved [ 46 , 26 ] . A leading hypothesis attributes them to fluctuations associated with a hidden liquid–liquid critical point (LLCP), buried below the freezing line at elevated pressure. Support for this view emerged from the observation of a first-order–like transition between two amorphous states in experiments [ 33 ] , and later between two liquid states in simulations [ 39 ] , giving rise to the liquid–liquid transition scenario. Experimental access to the LLCP has proven extremely challenging because it lies within the crystallization-prone no-man’s-land . Only very recent experiments have managed to penetrate this regime, probing various physical quantities on ultra-short time scales before ice formation intervenes [ 26 , 25 , 50 , 54 ] . As a consequence, much of the support for the LLCP hypothesis has relied on information coming from molecular dynamics simulations. Very recently, atomistic simulations using machine-learning based potentials trained on first-principles data [ 18 , 41 ] seem to confirm the original hypothesis of the existence of an LLT that were based on the ST2 potential [ 39 ] , while studies based on the classical point-charge model TIP4P/2005 have produced conflicting conclusions on the presence of an LLT [ 12 , 23 , 35 ] . Although these simulations have pushed the boundary to microsecond timescales, comparable to the ultrafast experiments that briefly access deeply supercooled water before crystallization intervenes, equilibration at these low temperatures becomes exceedingly challenging due to the very long structural relaxation times. At these conditions, the dynamics becomes glassy and the inference of thermodynamic quantities from simulations becomes difficult. This challenge also translates to experiments deriving thermodynamic support for the presence of an LLCP [ 54 ] , since they involve measurements limited to the microsecond timescales before the onset of crystallization. In fact, one of the open questions in the physics of supercooled water is the exact location of its glass transition temperature [ 2 ] . The commonly cited value of 136 K hinges on experiments of hyper-quenched and vapor-deposited samples whose relationship to slowly-supercooled pure water is uncertain. These measurements show only a faint calorimetric signal at this temperature, small enough to be interpreted as a ”shadow” transition [ 55 , 20 ] , suggesting that the true liquid–glass transformation may occur at substantially higher temperatures. While moderately concentrated salt solutions often exhibit a T_{g} around 136 K, recent experiments using minimal solute content find a markedly elevated T_{g}=166\pm 15 K [ 30 ] . Furthermore, salts believed not to perturb water structure yield a T_{g} near 190 K [ 58 , 52 ] . Studies in nanoconfinement—where crystallization is suppressed by geometry—likewise show that the 136 K feature is a shadow glass transition and place the genuine T_{g} between 170 and 200 K [ 32 ] . Temperature dependent electron diffraction experiments cooling from the liquid have shown that the structure does not evolve further below 200 K, placing the glass transition at this temperature for the cooling time of 15 \mu s before measurement [ 28 ] . Clearly, there is currently no smoking-gun measurement that unequivocally pinpoints the T_{g} of slow-cooled bulk water. Most importantly, if water’s T_{g} indeed lies at the upper end of the range suggested by the aforementioned experiments, it would be very close to the LLT temperature found in the most realistic ab initio–based water models based both on DFT [ 57 ] and MB-pol [ 5 ] , corrected for their melting point mismatch with respect to experiment. These observations raise an intriguing possibility: if deeply supercooled water becomes glassy around the temperature of the LLCP, could signatures often attributed to a liquid–liquid transition also be consistent with a transition between a fluid high-density liquid (HDL) and a kinetically arrested low-density glass? To tackle this question, we explore how the equilibrium dynamics of deeply supercooled water evolve as a function of temperature and pressure, comparing state-of-the-art machine-learning–accelerated first-principles simulations with experimental measurements in regimes where crystallization can be suppressed. This allows us to identify the conditions under which liquid water remains ergodic and when it instead becomes kinetically arrested—distinguishing genuine liquid behavior from glassy non-equilibrium states. By establishing the location of the glass-transition boundary relative to the proposed liquid–liquid transition, we suggest that signatures attributed to an LLT may not necessarily reflect an equilibrium transformation between two liquids, but could instead be consistent with a crossover between a liquid and a glass. Fluctuations between a liquid and an arrested state To appreciate the extent of the glassy dynamics that develops during the formation of the purported low-density liquid (LDL) phase, we examine simulation trajectories from the literature where LDL \leftrightarrow HDL fluctuations were previously reported, monitoring molecular motion and energy to distinguish liquid behavior from glassy arrest. Figure 1 presents three representative trajectories from two state-of-the-art deep neural-network (DNN) potentials, DNN@SCAN [ 18 ] and DNN@MB-pol [ 41 ] and the classical TIP4P/2005 model [ 12 ] . All temperatures reported for the DNN models are corrected for their melting point mismatch with respect to experiment, to allow comparison with experimental data. The transformation between the HDL and LDL phases are reflected in fluctuations observed in the potential energy of the system illustrated in the top panel for the three models on the microsecond timescale. To access the collective dynamics of system, we determined the time resolved mean square displacement (trMSD) defined as, \text{trMSD}(t)=\frac{1}{N}\sum\limits_{i=1}^{N}|\mathbf{r}_{i}(t)-\mathbf{r}_{i}(0)|^{2}, (1) which reports on the distance traveled by the molecules as a function of time. A striking feature observed from this analysis is the presence of long-lived periods where the trMSD remains essentially flat. These regions are color coded in light blue for both the potential energy and the trMSD which show that in the lower potential energy phase, the trMSD is mostly flat. This indicates that during these periods, the system is stuck in a dynamically arrested, low-density state due to vanishingly small molecular motion, whereas the high-energy

Read the complete text