Researchers at the University of Osaka have developed an artificial intelligence framework that evaluates how scientists describe the molecular structure of supercooled water, offering new insights into one of chemistry's most puzzling substances. By using a neural network to compare different structural descriptors, the AI helps identify which methods best capture the microscopic organization of water molecules as they shift between liquid states. The findings could improve our understanding of water's unusual physical properties.
Water behaves differently from most liquids, expanding as it freezes and exhibiting several other anomalies that become even more pronounced when it is supercooled. Scientists believe these behaviors arise from the competition between two liquid structures—a high-density liquid (HDL) and a low-density liquid (LDL). However, comparing the many proposed ways of describing these structures has long been a challenge because each framework measures molecular order differently.
To address this, the researchers trained an AI model to predict temperature from the molecular arrangement of water, allowing them to evaluate how effectively each structural descriptor represents the underlying physics. The neural network provided a unified and objective way to compare competing frameworks, revealing which descriptors contain the most meaningful information about water's molecular organization.
The study demonstrates how AI can accelerate discoveries in fundamental chemistry by helping researchers interpret complex molecular data. Beyond improving our understanding of water, the framework could be adapted to study other liquids and materials, advancing research in physics, chemistry, and materials science while supporting the development of more accurate computational models.
