Thursday 20 March 2025
Hydrogen, the simplest element in the universe, has long been a mystery. Despite its simplicity, it exhibits some surprisingly complex behaviors under extreme conditions – like high pressure. Scientists have been trying to understand these phenomena for decades, but progress has been slow due to the sheer computational power required.
Recently, a team of researchers used machine learning to crack the code on hydrogen’s liquid-liquid transition (LLT) at high pressures. The LLT is a phase transition where hydrogen transforms from a molecular crystal to an atomic liquid, which has important implications for our understanding of planetary science and materials modeling.
The team developed a fast and accurate machine-learning interatomic potential (MLIP) using a neural network architecture called MACE. They trained the model on density functional theory calculations and used it to drive classical and path-integral molecular dynamics simulations.
The results were striking. The MLIP predicted that the LLT is always supercritical above the melting temperature, meaning that it’s not a phase transition that occurs at a specific pressure, but rather a gradual change in behavior as the system is compressed. This challenges previous theories that suggested a first-order transition at high pressures.
To validate their findings, the researchers compared their results to ab initio molecular dynamics (AIMD) simulations. AIMD is a highly accurate method that uses quantum mechanics to simulate the behavior of molecules, but it’s computationally expensive and can only be used for small systems.
The comparison showed remarkable agreement between the MLIP and AIMD results. The MLIP accurately predicted the structural properties of hydrogen at high pressures and temperatures, including the radial distribution function and the molecular fraction.
The implications of this work are far-reaching. By understanding the LLT in hydrogen, scientists can gain insights into the behavior of other materials under extreme conditions. This knowledge can be used to improve our understanding of planetary science, particularly the formation and evolution of giant gas planets like Jupiter.
Moreover, the development of an accurate MLIP for hydrogen opens up new avenues for research on high-pressure materials. By applying machine learning to other systems, scientists may be able to uncover new phenomena that were previously inaccessible due to computational limitations.
Overall, this work demonstrates the power of machine learning in solving complex problems in physics. By combining advanced algorithms with cutting-edge computational resources, researchers can make significant progress in understanding some of the most challenging phenomena in the universe.
Cite this article: “Cracking the Code on Hydrogens Liquid-Liquid Transition”, The Science Archive, 2025.
Hydrogen, Machine Learning, High Pressure, Liquid-Liquid Transition, Interatomic Potential, Neural Network, Planetary Science, Materials Modeling, Molecular Dynamics, Quantum Mechanics.
