LLNL scientists use machine learning to study the behavior of “superionic water” in ice giants

According to foreign media reports, the internal water content of Uranus and Neptune is about 50000 times that of Earth's oceans, and a type of water called "superionic water" is believed to exist stably at depths exceeding the radius of Uranus and Neptune by about one-third. Superionic water is a stage of water in which hydrogen atoms become liquid while oxygen atoms remain solid on the lattice. Although the superionic state was proposed more than 30 years ago, its optical properties and oxygen lattice were only recently accurately measured in experiments by Marius Millot and Federica Coppari at the Lawrence Livermore National Laboratory (LLNL), and many of the characteristics of this thermal "black ice" are still unknown.

LLNL scientists have developed a new method to study the phase behavior of "superionic water" using machine learning at unprecedented resolution. Scientists say that most of the water buried deep in the planet's core in the universe may be "superionic water.". Understanding its thermodynamic and transport properties is crucial for planetary science, but it is difficult to detect through experiments or theory.

First principles molecular dynamics (FPMD) simulations predict that most of this water is in a superionic state under the pressure and temperature found in ice giant planets. However, this type of quantum mechanical simulation is traditionally limited to shorter simulation times (10 picoseconds) and smaller system scales (over 100 atoms), resulting in significant uncertainty in the position of phase boundaries, such as melting lines.

In the experiment of superionic water, sample preparation is very challenging: the position of hydrogen cannot be determined, and temperature measurement in dynamic compression experiments is not direct. Usually, experiments benefit from the guidance provided by quantum molecular dynamics simulations during the design phase and interpretation of results.

In recent research, the team has made a leap in their ability to handle large-scale and long-term systems by utilizing machine learning techniques to learn atomic interactions from quantum mechanical calculations. Then, they utilized the potential of machine learning to drive molecular dynamics and enabled cutting-edge free energy sampling methods to accurately determine phase boundaries.

LLNL physicist Sebastien Hamel said, "We use machine learning and free energy methods to overcome the limitations of quantum mechanical simulations and describe hydrogen diffusion, superion transition, and phase behavior of water under extreme conditions in water." He is a co-author of a paper published in Nature Physics.

The research team found that a phase boundary consistent with existing experimental observations helps to address the proportion of insulating ice, different superionic phases, and liquid water inside ice giants.

Building effective interaction potentials and maintaining the accuracy of quantum mechanical calculations is a challenging task. The framework developed here is universal and can be used to discover or describe other complex materials, such as battery electrolytes, plastics, nanocrystalline diamonds used in inertial confinement fusion (ICF) capsules, as well as new phases of ammonia, salts, hydrocarbons, silicates, and related mixtures related to planetary science.

Hamel said, "Our quantitative understanding of superionic water provides insights into the internal structure, evolution, and magnetic field of planets such as Uranus and Neptune, as well as the increasing number of icy exoplanets."

Prev: Scientists Successfully Constructed an Intelligent RNA Structure Prediction Model