A glass of water may look perfectly uniform, but at the molecular level, it could be carrying two different forms that are constantly swapping places. 

A close-up shot of water droplets. Image credits: Pixabay/Pexels

Scientists have suspected that liquid water is not a single microscopic structure but is made of two distinct local arrangements—a denser, more disordered one and a less dense, more ordered one. These structures continuously transform into one another. 

“According to the two-state hypothesis, liquid water can be viewed as a mixture of two distinct structures, A and B, but no one has ever seen a genuine ‘pure A’ or ‘pure B’ liquid water,” Xiao Cheng Zeng, a professor of material science and engineering at City University of Hong Kong, told Phys.org.

In their latest study, Zeng and his team used artificial intelligence to uncover what they say is the clearest molecular-level evidence yet for the existence of these two structures.

This finding could help explain some of water’s most puzzling behaviors and strengthen a long-contested theory about how liquid water is organized at the molecular level.

One reason water’s molecular structure has been a mystery is that researchers largely relied on measurements such as local density and molecular energy to identify different forms of water. 

While useful, these quantities could not clearly separate the two proposed structures, leaving the debate unresolved even after years of simulations and experiments.

The proposed transition between the two forms—known as a liquid-liquid phase transition (LLPT)—is believed to occur in deeply supercooled water, a state that is extremely difficult to study because water rapidly crystallizes into ice before scientists can observe it. 

Although experiments and computer simulations have hinted at the existence of this transition, they could not clearly identify the two molecular structures themselves. Even conventional simulation techniques that examined local density or molecular energy failed to separate them cleanly. 

Last year, another study attempted to narrow down the location of water’s proposed second critical point using deep neural networks. It came close but couldn’t directly reveal the two local structures. 

What changed this time

This time, instead of telling AI what to look for, the researchers let it discover the answer. So rather than designing rules for identifying the two structures, the team adopted an unsupervised deep learning approach. 

For instance, unlike conventional AI systems that learn from labeled examples, the researchers’ AI searches for hidden patterns without being told what the final answer should be. 

To give the model enough information, the researchers first carried out massive computer simulations using the widely used TIP4P/Ice model of water, which is designed to realistically reproduce how water molecules behave.

They generated about 74 million local molecular configurations, representing how individual water molecules and their neighbors were arranged across a wide range of temperatures and pressures. 

Around 17 percent of this data came from conditions close to the suspected liquid-liquid transition. The remaining came from outside that region, so the AI would learn water’s broader behavior instead of focusing on one special case.

Decoding water the AI way

The AI itself was built as an autoencoder, a neural network designed to compress complex information into a simpler internal representation before reconstructing it. First, the encoder analyzed the local environment around each water molecule while being trained to predict two familiar physical properties.

The first was how tightly packed nearby molecules were (local density), and the second was how strongly they interacted with one another (local potential energy). 

In this process, the encoder learned additional hidden characteristics that could distinguish water molecules in ways that density and energy alone could not. The decoder then used these hidden characteristics to reconstruct the original molecular structures.

More importantly, the researchers did not force the AI toward any preconceived picture of water. Instead, they applied only two loose mathematical constraints that controlled how strongly these hidden variables could relate to density and the geometric orientation of that relationship. 

The researchers then adjusted these constraints in many different ways until the hidden molecular patterns became clear.

Interestingly, the AI uncovered hidden structural characteristics that were largely independent of density. This allowed it to distinguish molecular arrangements that had remained invisible to traditional approaches focused mainly on density and energy.

When the researchers reached the optimal configuration, the AI separated the molecular data into two distinct clusters.

Plus, the system also identified a set of multidimensional reaction coordinates—mathematical variables that describe how a water molecule‘s local structure moves between the two states. These coordinates allowed the researchers to map the molecular transformation in far greater detail than had previously been possible.

One cluster corresponded to Structure A, where water molecules are packed more closely together in a denser, more disordered arrangement. The other represented Structure B, which is less dense and more ordered. 

“These findings provide molecular-level evidence in support of the two-state water model and may offer physical insights into the origin of the liquid–liquid phase transitions more generally,” the researchers note.

The researchers found these two local structures across a broad range of temperatures and pressures, including conditions approaching room temperature, suggesting they are a general feature of liquid water rather than something limited to extreme supercooled conditions.

Even more surprising, “the transformation between Structure A and Structure B is not a simple ‘back-and-forth’ process. The interconversion pathways of the two structures are different under different states of water,” Zeng added.

In the high-density liquid, the conversion follows one “upper semi-loop” route, crossing a single transition state. In the low-density liquid, it takes a different “lower semi-loop” route, also involving one transition state. 

However, near the boundary where the two liquid forms compete most strongly, these pathways merge into a much more complex three-dimensional full-loop pathway that passes through three separate transition states before completing the transformation. 

As water moves away from this boundary and one liquid state becomes dominant, the complex loop collapses back into the simpler semi-loop pathway.

An easy way to picture this is to imagine two hiking trails leading over a mountain. Most of the time, hikers take one direct route over a single pass. However, near a special region of the landscape, a complete circular trail opens up, allowing hikers to travel around the mountain before reaching the other side. 

According to the researchers, water molecules behave in a similar way, following different microscopic routes depending on where they are in the phase diagram.

The next step is experimental verification

The findings provide strong support for the idea that liquid water consists of two interconverting local structures, offering a potential explanation for some of water’s unusual properties, such as why water reaches its maximum density at 4°C (and not 0°C) and behaves unusually under pressure.

However, the work is based on computer simulations, and the hidden variables identified by the AI still need a clear physical interpretation and experimental verification.

The researchers now plan to determine exactly what these hidden characteristics represent and test whether they can be observed experimentally. 

“A crucial immediate step is to decode the physical interpretability of the two hidden physical characteristics revealed by the AI, and to seek their experimental verification,” Zeng said.

If confirmed, the results could deepen scientists’ understanding of water in biological systems, geological environments, and other settings where its unusual behavior plays an important role.

The study is published in the journal Nature Physics.