Water covers most of Earth’s surface, yet it behaves in ways that set it apart from nearly every other liquid. One of its most unusual traits is that it expands instead of contracts when it freezes. Scientists have long linked these odd behaviors to changes in water’s microscopic structure as temperature and pressure vary, but they have lacked a consistent way to describe and compare those structural changes.
Now, researchers at the University of Osaka have turned to artificial intelligence (AI) to tackle that challenge. Their AI system provides a unified way to compare different methods of describing the structure of supercooled water, helping identify which ones capture the most important features. The research was published in Communications Chemistry.
Why Supercooled Water Behaves So Strangely
For liquid water to become ice, its molecules must arrange themselves into an orderly crystal lattice. That process begins at a nucleation site, a surface where ice crystals can start forming. Tiny impurities in the water or even microscopic scratches inside a container can provide those starting points.
If those nucleation sites are absent, water can remain liquid even after it has been cooled below its normal freezing point. This unusual state is known as supercooled water.
Water’s unusual properties become even more pronounced under these conditions. Scientists believe these behaviors are linked to a balance between two competing forms of liquid water: a high density liquid (HDL) and a low density liquid (LDL). At the molecular level, water molecules are constantly forming and breaking networks of hydrogen bonds. As the temperature rises, the more compact HDL structures become increasingly dominant over the more open LDL arrangements.
AI Compares Competing Models of Water
Over the years, researchers have proposed many different ways to describe the local arrangement of water molecules, including measurements such as tetrahedral bond order and local density. Because these structural descriptors were developed independently, they use different scales, dimensions, and types of information. That has made it difficult to directly compare them and determine which are the most useful.
“Past studies have shown that using machine learning to classify and understand structural data is effective,” explains corresponding author Kang Kim. “We specifically wanted to incorporate a neural network model into this study to evaluate how accurate the descriptors were at capturing key structural information, in a way that is like human cognition.”
To train the AI, the researchers fed the neural network structural data generated from molecular dynamics simulations of supercooled water. Through repeated trial and error, the system learned to recognize meaningful patterns in the molecular structures.
New Clues to Water’s Hidden Structure
“The network used what it had learned to compare how 16 descriptors differentiated between LDL and HDL structures at different temperatures,” reports Nobuyuki Matubayasi, senior author. “In this way, we determined the most efficient descriptors.”
The researchers say their framework could improve scientists’ understanding of how microscopic structural changes are connected to the thermodynamic behavior of water. The findings may also help explain the origin of water’s unusual properties while guiding the development of even better tools for studying its complex molecular structure.