Machine Learning Helps Create Detailed, Efficient Models of Water

Models use a fraction of the computational cost of today’s best atom-based water models.

Four blue cube graphs depicting thermodynamic water flow
Image courtesy of Subramanian Sankaranarayanan, Argonne National Laboratory
Time series of a machine-learning–based coarse-grained simulation, ML-BOPdih, provides snapshots spanning ~1 microsecond (t=time) showing evolution of grain boundaries (green) between regions of hexagonal (blue) and cubic (orange) ice.

The Science

How water acts affects everything from storm clouds to ice sheets. Computer scientists want to model water’s various properties. Accurate and computationally efficient molecular-level descriptions of large samples of ice-water systems are difficult to build. Why? The numerous molecules and various timescales remain a challenge despite advances in computing hardware. Now, a team developed machine-learning–based water models that correctly predict water’s key features, such as the melting point of ice. The team’s models use a fraction of the computational cost of the best atomistic water models available today.

The Impact

The team devised a way to better model water’s properties. They developed a machine-learning workflow. It offers accurate and computationally efficient models. Researchers can apply the A.I.-based approach to other material models to improve predictions.


The team developed three machine-learning–based (ML) coarse-grained water models that accurately describe the structure and thermodynamic anomalies of both water and ice at larger, mesoscopic scales based on molecular-level interactions. All three models are two orders of magnitude cheaper—in terms of computational cost—than today’s atomistic models. In these models, individual water molecules are modeled as single particles that interact. In a major departure from traditional approaches, the team trained the models against first-principles calculations, experimental results, and temperature-dependent properties from molecular simulations. To overcome optimization challenges, the team employed a multi-level hierarchical workflow using global and local optimization algorithms coupled with on-the-fly molecular simulations. They used the global algorithm (Genetic Algorithm) to generate a broad survey of the parameter landscape, which is followed by multiple refinements using a local optimization algorithm with tens of milliseconds of simulation data generated on-the-fly to compute temperature-dependent properties for candidate models.

Two models successfully predicted a range of thermodynamic properties, including the density maximum of liquid water, the melting transition temperature, and the self-diffusion coefficient (a dynamic property). The team used the models on Mira, a supercomputer at the Argonne Leadership Computing Facility, to investigate homogeneous nucleation of supercooled water leading up to the formation and growth of ice grains and made discoveries about the energy involved in forming ice. Also, the team examined how their machine-learning workflow could be used to improve existing water models.


Subramanian Sankaranarayanan
Center for Nanoscale Materials
Argonne National Laboratory


The Center for Nanoscale Materials was supported by the Department of Energy (DOE), Office of Science, Office of Basic Energy Sciences. This research used resources of two DOE Office of Science Advanced Scientific Computing Research user facilities: the Argonne Leadership Computing Facility and the National Energy Research Scientific Computing Center. The Carbon, Fusion, and Laboratory Computing Resource Center computing facilities at Argonne National Laboratory were also used.


H. Chan, M.J. Cherukara, B. Narayanan, T.D. Loeffler, C. Benmore, S.K. Gray, and S.K.R.S. Sankaranarayanan, “Machine learning coarse grained models for water.” Nature Communications 10, 379 (2019). [DOI: 10.1038/s41467-018-08222-6]

Highlight Categories

Program: ASCR , BES , SUF

Performer: University , DOE Laboratory , SC User Facilities , ASCR User Facilities , NERSC , ALCF , BES User Facilities , CNM