New Insights into Water’s Molecular Behavior

Scientists have long speculated that water behaves as if it consists of two distinct liquids—a denser one and a less-dense counterpart—constantly interchanging between the two. However, capturing concrete evidence of this molecular switch has proven challenging. Recently, researchers, aided by artificial intelligence, claim they have finally achieved this breakthrough.

“It’s quite perplexing—just one water, right?” remarked Xiao Cheng Zeng, a physical chemist at the City University of Hong Kong and co-author of the recent study. His curiosity led him to explore scientific literature, where he stumbled upon the two-state hypothesis. “That piqued my interest. We had plenty of discussions on the topic but lacking in solid proof,” he added.

The findings, which appeared on June 4 in Nature Physics, not only substantiate this elusive molecular transformation but may also clarify numerous peculiarities associated with water’s behavior.

Typically, liquids become denser when cooled, but water is an exception; it densifies until about 4 degrees Celsius and then begins to expand, which is why ice floats. Water also exhibits unique resistance to temperature fluctuations compared to other liquids and has a viscosity that diminishes under specific pressure conditions. Researchers suspect these various anomalies may share a common thread.

The two-state model aims to serve as that unifying theory.

A Long-held Suspicions

Zeng has been exploring the properties of water since his early research days in the late 1990s, focusing on its freezing behavior. He first encountered the two-state hypothesis around 2006 but felt it was too complex to pursue at that time. Things shifted around 2016, as experimental evidence emerged suggesting that supercooled water might separate into high-density and low-density forms.

About two and a half years ago, Zeng tasked postdoctoral researcher Liwen Li with the problem. Instead of revisiting traditional methods that others struggled with, Li proposed leveraging “unsupervised deep learning”—where AI identifies patterns in data on its own.

“With AI, it had to learn—utilizing its knowledge to analyze and discover,” Zeng explained.

The team conducted extensive molecular dynamics simulations using the GROMACS package, observing how vast numbers of water molecules interacted and generated millions of data points. “In the past, you’d need a bunch of students for that. With AI, it took Li maybe a year and a half. Without it, that might have stretched out to a decade,” Zeng noted.

Two Routes of Transformation

The researchers discovered that the conversion paths between the two structures varied under different conditions. Most often, the transformation occurs along what the team terms a “semi-loop” pathway, requiring just one energy barrier. However, near the boundary between high-density and low-density water (similar to where ice and liquid state coexist at freezing point), molecules could opt for a more complex “full-loop” route that involves three distinct barriers.

Zeng likened this scenario to hiking on a mountain divided in two: one gentle slope and one steep cliff. While most hikers would stick to the slope—the semi-loop—closer to the boundary, it’s as if hikers can encircle the peak, representing the full loop.

The team is now developing a more sophisticated machine-learning model to validate their findings, aspiring to correlate it with properties such as density, viscosity, and temperature.

Proving this structure in actual water will be challenging. Zeng anticipates that new experimental techniques—like those created at the Pacific Northwest National Laboratory, which previously provided indirect evidence of water’s two-state nature—will be necessary.

“Once we can validate this through experiments,” he remarked, “this model will enhance our understanding of how water interacts in nature.” With most biological and pharmaceutical processes occurring in water, a better grasp of its molecular structure could illuminate how various substances, like salts and proteins, behave in solution. While those insights are crucial for developing injectable drugs and maintaining cellular functions, practical applications are still a long way off.