AI accurately predicts temperature from molecular arrangements

A research group consisting of Doctoral Student Kohei Yoshikawa (first year), Doctoral Student Kokoro Shikata (third year), Associate Professor Kou Kim, and Professor Nobuyuki Matubayasi of the Graduate School of Engineering Science at the University of Osaka has discovered that by training artificial intelligence on hydrogen bond networks in liquid water, temperature can be predicted from the arrangement information of water molecules. The findings were published in Communications Chemistry.

In liquid-state water, adjacent water molecules constantly swap their hydrogen-bonding partners while moving in a cooperative manner. Because the hydrogen bond network continuously rearranges over time, it is not easy to find clear order or regularity. On the other hand, since water molecules have a bent structure and exhibit strong polarity, hydrogen bonds tend to locally form tetrahedral structures. It has been thought that remnants of this structure exist even in the liquid state.

The research group clarified that an artificial intelligence using a neural network can predict temperature through the structures found in molecular arrangement information of water under different temperature conditions generated by molecular dynamics simulations. Furthermore, they systematically compared and evaluated the temperature prediction performance of 16 types of molecular-level structural indices that have been proposed so far.

As a result, they found that an index quantifying the structural disorder caused by a fifth water molecule approaching the four water molecules forming a tetrahedral structure showed the highest prediction accuracy. This result suggests that while local tetrahedral order inherently exists even in liquid-state water, that order is constantly perturbed by interactions with surrounding water molecules.

In other words, this result can be interpreted as showing that the tetrahedral structure is not completely collapsed. Rather, liquid-state water exists in a state where "partial order" and "dynamic disorder" coexist. Additionally, from analysis of temperature dependence, it was clarified that as the temperature decreases, this local tetrahedral order develops more prominently, and the structure continuously changes to a more ordered state. On the other hand, it was also confirmed that the order still involves dynamic disorder, maintaining properties unique to liquids that are different from a perfect crystal structure.

In this study, the group showed that even in a substance that appears disordered at first glance, such as liquid water, structural information reflecting thermodynamic states such as temperature is included. This provides a new perspective of deciphering the properties of substances from molecular-level structural fluctuations. This approach could lead to clarifying the origin of water's anomalous properties, such as its density peaking at 4℃.

Kim stated, "The structure of liquid-state water is highly disordered, and we initially did not imagine that temperature could be predicted solely from the arrangement of water molecules. However, the neural network captured the information hidden within it and deciphered the temperature with high accuracy. Furthermore, we were surprised to find a molecular-level basis for this high accuracy in the disorder of the tetrahedral structure. We hope this achievement will serve as an opportunity to deepen our understanding of water, which is the most familiar substance to us."

Journal Information
Publication: Communications Chemistry
Title: Machine learning evaluation of structural descriptors for supercooled water
DOI: 10.1038/s42004-026-02097-1