Water covers much of the Earth’s surface, but it behaves differently than almost all other liquids. One of its most unusual characteristics is that it expands rather than shrinks when frozen. Scientists have long linked these strange behaviors to changes in water’s microstructure as temperature and pressure change, but lacked a consistent way to explain and compare these structural changes.
Now researchers at Osaka University are turning to artificial intelligence (AI) to tackle this challenge. The company’s AI system provides a unified way to compare different methods of describing the structure of supercooled water and helps identify which methods capture the most important features. This study communication chemistry.
Why does supercooled water behave strangely?
For liquid water to turn into ice, its molecules must be arranged in an ordered crystal lattice. The process begins at the nucleation site, the surface where ice crystals begin to form. Small impurities in the water or even microscopic scratches in the container can be their starting point.
If these nucleation sites were not present, water could remain liquid even after cooling below its normal freezing point. This abnormal condition is known as supercooled water.
Under such conditions, the unusual properties of water become even more pronounced. Scientists believe that these behaviors are related to the balance between two competing forms of liquid water: high-density liquid (HDL) and low-density liquid (LDL). At the molecular level, water molecules are constantly forming and breaking networks of hydrogen bonds. As temperature increases, more compact HDL structures become predominant over more open LDL arrangements.
AI compares competing water models
Over the years, researchers have proposed various methods to describe the local arrangement of water molecules, including measurements such as tetrahedral bond order and local density. These structural descriptors were developed independently and therefore use different scales, dimensions, and types of information. This makes it difficult to directly compare them and determine which is most useful.
“Past research has shown that using machine learning to classify and understand structured data is effective,” explains corresponding author Kang Kim. “We specifically wanted to incorporate neural network models into this study to assess how accurately the descriptors could capture important structural information in a manner similar to human cognition.”
To train the AI, the researchers fed a neural network with structural data generated from molecular dynamics simulations of supercooled water. Through repeated trial and error, the system was able to recognize meaningful patterns in molecular structure.
New clues to the hidden structure of water
“The network used what it learned to compare how the 16 descriptors discriminate between LDL and HDL structures at different temperatures,” reports senior author Nobuyuki Matsubayashi. “In this way, we determined the most efficient descriptor.”
The researchers say their framework can improve scientists’ understanding of how microscopic structural changes relate to water’s thermodynamic behavior. The discovery could help explain the origin of water’s unusual properties and help develop even better tools to study water’s complex molecular structure.
