Researchers at UC San Diego have produced what may be the strongest computational evidence yet that liquid water can split into two distinct forms under extreme cold and high pressure. Their findings, reported in a recent study, locate a hidden critical point near 200 K and roughly 1,250 atmospheres, deep in a temperature zone where water resists crystallization long enough for two separate liquid phases to emerge. The result strengthens a hypothesis that has divided physicists for more than a century and could reshape how scientists understand water’s many unusual behaviors.
Simulating Water Where Experiments Cannot Reach
Water does not always freeze when cooled below zero degrees Celsius. Under the right conditions, it can remain liquid well below its normal freezing point, entering a supercooled state. The catch is that this deeply supercooled zone, sometimes called “no-man’s land,” sits at temperatures so low and pressures so high that bulk water crystallizes almost instantly, making direct laboratory observation extraordinarily difficult.
To get around that barrier, the UC San Diego team ran microsecond-scale molecular dynamics simulations using a neural-network potential trained on the well-regarded MB-pol water model. This deep-learning framework, known as DNN@MB-pol, was first developed and validated in work published in Nature Communications, which showed that it can reproduce an accurate phase diagram when nuclear quantum effects are included. By scaling that framework to much longer simulation times and larger systems, the researchers could watch supercooled water evolve over timescales that previous models could not reach.
The simulations revealed two discernibly different liquid states coexisting in the deeply supercooled, high-pressure regime. One form is denser and less ordered; the other is lighter and more structured. Their coexistence mirrors a phenomenon already known in amorphous (non-crystalline) ice, where high-density amorphous and low-density amorphous forms have been documented for decades. Finding a liquid analog of that duality has been the central goal of this line of research.
By systematically varying temperature and pressure, the team traced a boundary where the two liquid phases can coexist and identified a critical point where that boundary ends. Near this point, small fluctuations in density and local structure become enormous, giving rise to the dramatic changes in response functions, such as compressibility and heat capacity, that have long puzzled water researchers. The simulations suggest that the critical point lies squarely inside no-man’s land, explaining why it has eluded direct experimental detection.
A Century-Old Puzzle About Water’s Odd Behavior
The debate over whether water harbors a second liquid phase stretches back more than a hundred years. Water is one of the few substances that can exist in nature as a solid, liquid and gas under everyday conditions, and it displays a long list of properties that defy simple explanation: it reaches maximum density at 4 degrees Celsius, ice floats rather than sinks, and its heat capacity is anomalously high. A liquid-liquid critical point, if it exists, would provide a single theoretical origin for many of these quirks.
Researchers at Stockholm University and elsewhere have argued that such a critical point would be the source of water’s strange properties, a view echoed in related commentary on the new simulations. Earlier computational work helped build the case. A study from Princeton provided evidence that one liquid state of water is more locally ordered than the other, with a more tetrahedral hydrogen-bond network. Separate simulations reported that the two liquid forms are “entangled” at the microscopic level, meaning their hydrogen-bond networks interweave in ways that make clean separation difficult to observe in short experiments.
On the experimental side, progress has been slower because of the rapid onset of crystallization. In 2020, X-ray laser experiments followed structural changes in supercooled water in real time, hinting at a transformation between two liquid-like arrangements but stopping short of pinning down a critical point. These efforts set the stage for the new computational work by clarifying what structural signatures to look for and which regions of the phase diagram are most promising.
Converging Evidence From Multiple Angles
What makes the new result significant is not just its own data but how it fits into a growing body of independent evidence. Experimental teams have attacked the problem from several directions. One group used ultrafast X-ray scattering at the LCLS free-electron laser to measure structural signatures of water below the homogeneous ice nucleation temperature, catching fleeting glimpses of the liquid before it crystallized. That work confirmed that structural changes consistent with two liquid forms do occur in no-man’s land, even if the window for observation is vanishingly brief.
Other researchers have tried to slow crystallization by working with confined or impure systems. A study in Science reported a liquid–liquid transition in a supercooled aqueous solution and connected it to the known high-density and low-density amorphous ice transition. By adding solutes, the authors could depress the freezing point and stretch out the time before crystallization, giving them a better chance to observe a transition between two liquid-like states.
Meanwhile, a separate computational study published in Physical Review Letters found evidence for a liquid–liquid phase transition in simulated supercooled water nanodroplets. Tiny droplets can resist crystallization longer than bulk samples because the formation of an ice nucleus is less favorable in confined geometries. The nanodroplet simulations therefore provide a bridge between idealized bulk models and the kinds of microscopic water samples that can be probed experimentally.
Each of these results, taken alone, leaves room for alternative explanations. Simulations depend on the accuracy of their underlying models. Experiments in solutions or nanodroplets introduce variables that might not apply to pure bulk water. But the UC San Diego work narrows the gap between simulation and experiment by using a potential that closely matches real water’s known phase behavior, then pushing it into the regime where the critical point should appear.
Why the Computational Method Matters
A common criticism of earlier simulation studies was that simplified water models might produce artifacts, meaning the two-liquid behavior could be a quirk of the model rather than a real physical phenomenon. The DNN@MB-pol approach addresses that concern directly. By training a deep neural network on high-accuracy quantum-mechanical data and validating it against experimental phase diagrams, the UC San Diego researchers built a model that reproduces water’s real-world thermodynamics with unusual fidelity.
The simulation campaign itself was massive, drawing on compute resources from systems such as Expanse at the San Diego Supercomputer Center and additional national facilities. Running microsecond-scale trajectories for thousands of water molecules at many different temperature-pressure combinations required petascale computing power. Those long trajectories were crucial: if simulations are too short, they may capture only transient fluctuations rather than true equilibrium behavior, making it easy to misinterpret rare density fluctuations as evidence for a separate phase.
By extending the simulations, the team could watch the system cross back and forth between low-density and high-density liquid states, map out coexistence lines and determine how key thermodynamic quantities diverge near the critical point. They also checked for finite-size effects by repeating calculations with different system sizes, ensuring that the observed two-phase behavior was not just a finite-box artifact.
Another important advance was the explicit treatment of nuclear quantum effects, which can subtly alter hydrogen bonding at low temperatures. In earlier generations of water models, these effects were either neglected or treated approximately, potentially shifting or even eliminating the predicted critical point. The MB-pol framework, and by extension DNN@MB-pol, incorporates these quantum contributions in a way that has been benchmarked against spectroscopy and thermodynamic data, lending additional weight to the new predictions.
Implications and Open Questions
If the liquid–liquid critical point described by the UC San Diego simulations is real, it offers a unifying explanation for many of water’s anomalies. Properties such as compressibility, thermal expansivity and heat capacity all show sharp changes as liquid water is cooled toward the supercooled regime. In the critical-point picture, these anomalies arise because the system is flirting with a hidden transition between two structurally distinct liquids, even if it never actually crosses that boundary under everyday conditions.
The work also has implications for fields ranging from cryobiology to planetary science. Supercooled water exists in high-altitude clouds and in the icy mantles of outer solar system bodies, where pressures and temperatures can approach the regime explored in the simulations. Understanding how water’s structure responds under those conditions could improve models of cloud formation, ice nucleation and even the internal dynamics of icy moons.
Still, the new findings do not close the book on the debate. Direct experimental confirmation of a bulk liquid–liquid critical point remains out of reach because of rapid crystallization in no-man’s land. Future experiments may exploit even faster X-ray probes, more extreme confinement or novel sample preparation techniques to push closer to the predicted critical region. On the theory side, independent simulations using different high-accuracy potentials will be needed to test the robustness of the DNN@MB-pol predictions.
For now, the UC San Diego study significantly shifts the balance of evidence. By combining a state-of-the-art machine-learning potential with large-scale computing, the researchers have drawn the sharpest picture yet of water’s hidden dual personality, suggesting that beneath the familiar liquid we drink lies a more complex landscape of structures and phases than meets the eye.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
