Artificial Intelligence Uncovers Two Hidden Molecular Structures in Water
Artificial Intelligence (AI) has been used to find the most powerful evidence at the molecular level for the idea that liquid water is made up of two different but constantly changing local structures. This is a long-standing idea known as the two-state model of water, and it could help explain many of the unusual properties of water.
Many researchers have thought that liquid water does not consist of only one type of molecular arrangement, but rather consists of two different local structures or arrangements in water. The first local structure, called HDLS (High Density Liquid Structures), is disordered, closely packed, non-tetrahedral, and is thought to be denser and less organized than the second local structure, called LDLS (Low Density Liquid Structures), which is ordered, open, tetrahedral, and is thought to be less dense and more organized. Scientists have not yet been able to distinguish these two molecular arrangements unambiguously.
One of the reasons that the two local structures of water were not able to be distinguished until now is that earlier studies were mostly based upon physical measurements such as local density and molecular energy. These types of measurements cannot be
used to unambiguously determine the difference between the two proposed arrangements. Also, most researchers believe that the transition between the two different local structures, which is called the Liquid-Liquid Phase Transition (LLPT), occurs in water when it is cooled below its normal freezing point without turning into ice and, therefore, cannot be studied easily.Researchers, directed primarily by Professor Xiao Cheng Zeng at City University of Hong Kong, opted for a different method of research compared with previous approaches for their most recent study. Researchers used an unsupervised deep learning model called an autoencoder. Unlike conventional Artificial Intelligence (AI) systems, it was not given predefined structural rules or labels. The autoencoder analysed the local molecular environment surrounding each water molecule and learned hidden (latent) structural features beyond conventional physical properties such as local density and local potential energy. This enabled it to classify local molecular environments directly from the simulated molecular data without relying on predefined structural rules, traditional structural measurements, or human-defined order parameters.
Prior to applying machine learning techniques, the researchers used MD (Molecular Dynamics) Simulations, using water models such as TIP4P/Ice. This is a rigid four-site point-charge water potential model and is widely used to reproduce the thermodynamic and structural properties of water and ice. This model can simulate approximately 74 million unique molecular configurations, which represent the spatial arrangement of water molecules and their neighbouring molecules, arranged at different temperatures and pressures.
Approximately 17% of these molecular configurations were sampled from conditions close to the proposed Liquid-Liquid Phase Transition (LLPT), while the remaining data represented liquid water across a much broader range of temperatures and pressures, including conditions close to ambient. This broader training dataset enabled the AI to learn the general structural behaviour of liquid water rather than focusing solely on the supercooled regime, demonstrating that the two local molecular environments are intrinsic characteristics of liquid water rather than features limited to the LLPT region.
This process allowed the researchers to identify and classify two general molecular structures based upon their behavior through thermodynamic processes: HDLS (High-Density Liquid Structures) and LDLS (Low-Density Liquid Structures), and then mapped how changes in temperature and pressure shift the balance between these two molecular environments that lead one molecular structure into another. Importantly, the dynamic thermodynamic equilibrium of molecule-based structure variability shown by the existence of two separate structure types changes, as does the configuration of the structures themselves, as they compete for existence based upon variances in temperature and/or pressure applied to water molecules.
The two local structures were found for a wide temperature and pressure range (from close to room temperature), showing that they are general features of all liquid water, not only limited to deeply supercooled water.
The authors state that these findings may help explain several anomalous properties of water, including why its density reaches a maximum at 4°C. As water warms from 0°C to 4°C, many of the open, tetrahedral Low-Density Local Structures (LDLS) collapse into more compact High-Density Local Structures (HDLS), causing the liquid to contract and its density to increase. Above 4°C, normal thermal expansion becomes dominant, causing the density to decrease. The researchers also suggest that the coexistence of HDLS and LDLS may explain water’s unusual behaviour under pressure. However, they emphasise that these findings are based on molecular simulations, and the AI-discovered structural features must be physically interpreted and validated through experimental studies before definitive conclusions can be drawn.
