What’s going on inside the most distant planets in our solar system? The ice giant planets Uranus, Neptune, and sub-Neptune exoplanets are believed to be composed primarily of the elements hydrogen, carbon, nitrogen, and oxygen (H-C-N-O). However, the details of their interior structures have remained a mystery due to unclarity around the behavior of compounds H2O, NH3, and CH4 at the high pressures and temperatures found inside these planets.

Recently, researchers at UC Berkeley used the Perlmutter supercomputer at the National Energy Research Scientific Computing Center (NERSC) to make progress toward a better understanding of H-C-N-O chemistry, a step forward for planetary science that may also have applications on Earth. Their work was published in Nature Communications in November.

Using Perlmutter, UC Berkeley Earth and Planetary Sciences Ph.D. student Kyla de Villa and her team from the Militzer Group produced ab initio simulations of 13 H-C-N-O compounds and demonstrated that at high temperatures, they become superionic – a phase of matter between solid and liquid in which light ions (hydrogen in this case) diffuse like a liquid through a solid sublattice of larger nuclei. At even higher temperatures, four of the 13 compounds demonstrated double superionicity: in addition to hydrogen, either carbon or nitrogen also diffused through the sublattice of the remaining large nuclei.

“This work has been very exciting because it has made us totally rethink what the behavior of planetary ices must be deep inside planets,” said de Villa. “It was previously shown that water and ammonia become superionic at high pressures, but here we see that superionic diffusion of light elements is exceedingly common at high enough temperatures and pressures. Our observation of double superionicity further challenges assumptions about how superionic diffusion occurs.”

Identifying superionicity and double superionicity in these compounds at high temperature and pressure expands scientists’ understanding of which compounds may appear in the form of hot ices in the mantles of ice giant planets. Furthermore, the superionicity of these compounds may be the driver of Uranus and Neptune’s unusual asymmetrical, non-dipolar magnetic fields.

“We predict that superionicity will be a common feature among compounds in the interiors of Uranus and Neptune,” said Burkhard Militzer, UC Berkeley professor and head of the Militzer Group, who is also an author on the paper. “It will contribute to electrical conductivity and thus enhance the generation of the magnetic fields in their interiors.”

Applications for superionic properties aren’t limited to planetary sciences, however. Superionic materials are ionically conductive due to the flow of charged particles, even at pressures and temperatures that are too low to allow them to be electronically conductive. Many solid-state batteries incorporate superionic materials, and a better understanding of superionicity could help scientists create better batteries for a variety of applications.

To simulate each compound, the researchers first performed density functional theory (DFT) calculations using the Vienna Ab Initio Simulation Package (VASP), a code for atomic-scale materials modeling. However, to work around the computational cost and limitations to size and timescale of DFT simulations and allow for scaled-up explorations, they went on to generate machine-learning potentials – functions that accurately express atomic structures through their potential energies in a way that is less computationally expensive. They produced the potentials by running a large number of DFT calculations and feeding the results into a neural network, which generated a potential energy surface that could then be used to run much faster and less expensive simulations using the LAMMPS molecular dynamics simulator.

Ultimately, this method provided significant improvements in speed and efficiency over previous methods. The researchers were able to perform AI-accelerated molecular dynamics simulations of a sample of 1,925 atoms on using Perlmutter in 1/25 the time required to simulate 96 atoms on twice as many nodes using VASP. Additionally, Perlmutter provided a much faster simulation experience than Cor, NERSC’s previous systemi: the researchers saw a 6x speedup on Perlmutter for VASP simulations, a 30x speedup for training the neural network, and more than a 100x speedup for molecular dynamics simulations using LAMMPS.

“This project would not have been possible without the resources of NERSC,” said Felipe González, another author on the paper. “Running long, expensive simulations with VASP for training and subsequent machine learning with LAMMPS for deployment required massive parallelism in both codes and quick access to these resources, which the Perlmutter cluster gave us.”