A cost-effective machine learning potential enabled the investigation of hydrogen phase transitions at temperatures between 100 and 4000 K and pressures between 25 and 400 gigapascals with converged simulation size and time. Photo credit: @Michele Ceriotti, EPFL
The hydrogen atomWith its single proton orbited by a single electron, it’s arguably the simplest material out there. Elemental hydrogen can nonetheless show an extremely complex behavior – for example, at megabar pressures it changes from an insulating liquid to a metallic conductive liquid.
While the transition is simply fascinating from the point of view of the physics and materials science of condensed matter – liquid-liquid phase transitions are rather unusual – it also has significant implications for planetary research, as liquid hydrogen can be found in the interior of giant planets like e.g. Jupiter and Saturn as well as brown dwarf stars. Understanding the liquid-liquid transition is then a central part of accurately modeling the structure and evolution of such planets, and standard models generally assume a sharp transition between the insulating molecular fluid and the conductive metallic fluid. This sharp transition is associated with a discontinuity in density and therefore a clear boundary between an inner metal jacket and an outer insulating jacket in these planets.
While scientists have put significant effort into studying and characterizing this transition, as well as the many unusual properties of dense hydrogen – including the rich and poorly understood solid polymorphism, the anomalous melt line, and the possible transition to a superconducting state – the laboratory investigation is due complicated by the need to create a controllable high pressure and temperature environment and to include hydrogen during measurements. Experimental research has not yet reached a consensus on whether the transition is abrupt or smooth, and various experiments have located the liquid-liquid transition at pressures that are up to 100 gigapascals apart.
“The kind of experiment you need to do to examine a material with the same pressure range that you find on Jupiter is most trivial,” Ceriotti said. “Because of the limitations, a lot of different experiments were run and the results were very different from one another.”
Although the modeling techniques introduced over the past decade have enabled scientists to better understand the system, the tremendous computational effort involved in solving the essential quantum mechanical problem for the behavior of hydrogen atoms has resulted in these simulations necessarily being timed to one The scale was limited to a few picoseconds and only a few hundred atoms. The results here are also mixed.
To investigate the problem more closely, Ceriotti and colleagues Bingqing Chen from the University of Cambridge and Guglielmo Mazzola from IBM Research Zurich used an artificial neural network architecture to build up a machine learning potential. Based on a small number of very precise (and time-consuming) calculations of the problem of the electronic structure, the inexpensive machine learning potential made it possible to investigate hydrogen phase transitions at temperatures between 100 and 4000 K and pressures between 25 and 400 gigapascals. with converged simulation size and time. The simulations, which are carried out mainly on EPFL computers at SCITAS, took only a few weeks compared to the 100 million years of CPU time that would have been required to carry out conventional simulations to solve the quantum mechanical problem.
The resulting theoretical study of the phase diagram of dense hydrogen allowed the team to reproduce the re-entry melt behavior and polymorphism of the solid phase. Simulations based on the machine learning potential showed, contrary to the general assumption that hydrogen undergoes a first-order phase transition, indications of continuous metallization in the liquid. This in turn not only points to a smooth transition between insulating and metallic layers in giant gas planets, but also compensates for existing discrepancies between laboratory and modeling experiments.
“When high pressure hydrogen is supercritical, as our simulations show, there is no sharp transition where all properties of the liquid suddenly jump,” said Ceriotti. “Depending on the exact property you are testing and the way you define a threshold, the transition will take place at a different temperature or pressure. This could reconcile a decade of controversial results from high pressure experiments. Different experiments measured slightly different things and failed to identify the transition at the same point because there is no sharp transition. ”
To reconcile their results with some earlier models that actually identified a sharp transition, Ceriotti said that only in small simulations they could see a marked jump in properties and that in those cases they were more likely than able to follow the jump to solidification to a liquid-liquid transition. The observed sharp transition should then be understood more as an artifact of the limitations in using simulations based on traditional physics-based modeling. The machine learning approach has enabled researchers to run simulations that are typically four to ten times larger and several hundred times longer. This gives them a much better view of the whole process.
While it was applied to a topic related to planetary science in this particular article, the same technology can be applied to any problem in materials science or chemistry, Ceriotti said.
“This is a demonstration of a technology that allows simulations to get into a regime that was previously impossible,” said Ceriotti. “The same technology that could help us better understand the behavior of planets can also be used to develop better drugs or better performing materials. There really is the potential for simulation-driven change in the way we understand the behavior of everyday and exotic matter. ”
Reference: “Notes on the supercritical behavior of liquid high pressure hydrogen” by Bingqing Cheng, Guglielmo Mazzola, Chris J. Pickard and Michele Ceriotti, September 9, 2020, nature.
DOI: 10.1038 / s41586-020-2677-y
