Researchers at the University of Osaka have developed an AI-driven framework to analyze the molecular structure of supercooled water, a state where water remains liquid despite being cooled below its normal freezing temperature. The study, published on July 6, 2026, in Communications Chemistry, utilizes machine learning to evaluate how different structural descriptors capture the transition between high-density liquid (HDL) and low-density liquid (LDL) states. These two structures arise from competing hydrogen bond networks within water, which researchers believe underpin the substance's unique physical anomalies, such as its expansion upon freezing.

To implement this, the research team used a neural network model to mirror human cognitive processes in recognizing structural patterns. They trained the system using structural data derived from molecular dynamics simulations. The AI tested and compared 16 distinct structural descriptors—including measurements like tetrahedral bond order and local density—to determine their individual efficiency in differentiating between HDL and LDL configurations. According to the study, the AI identified which descriptors most effectively represent the microscopic changes connected to water's thermodynamic behavior.

The findings, led by corresponding author Kang Kim and senior author Nobuyuki Matubayasi, aim to provide a more consistent method for measuring structural shifts in water. By evaluating these descriptors on a unified scale, the scientists established a tool to correlate microscopic molecular arrangements with macroscopic physical properties. This approach may help researchers better understand why water behaves differently from most other liquids and refine future methodologies for examining water's fundamental nature and behavior across varying temperature and pressure conditions.

Researchers at the University of Osaka have developed an AI-driven framework to analyze the molecular structure of supercooled water, a state where water remains liquid despite being cooled below its normal freezing temperature. The study, published on July 6, 2026, in Communications Chemistry, utilizes machine learning to evaluate how different structural descriptors capture the transition between high-density liquid (HDL) and low-density liquid (LDL) states. These two structures arise from competing hydrogen bond networks within water, which researchers believe underpin the substance's unique physical anomalies, such as its expansion upon freezing.

To implement this, the research team used a neural network model to mirror human cognitive processes in recognizing structural patterns. They trained the system using structural data derived from molecular dynamics simulations. The AI tested and compared 16 distinct structural descriptors—including measurements like tetrahedral bond order and local density—to determine their individual efficiency in differentiating between HDL and LDL configurations. According to the study, the AI identified which descriptors most effectively represent the microscopic changes connected to water's thermodynamic behavior.

The findings, led by corresponding author Kang Kim and senior author Nobuyuki Matubayasi, aim to provide a more consistent method for measuring structural shifts in water. By evaluating these descriptors on a unified scale, the scientists established a tool to correlate microscopic molecular arrangements with macroscopic physical properties. This approach may help researchers better understand why water behaves differently from most other liquids and refine future methodologies for examining water's fundamental nature and behavior across varying temperature and pressure conditions.