Exchange unit: Cambridge University

DOI: 10. 1038/s 4 1586-020-2677-y

Hydrogen is the simplest and most abundant element in the universe, and it will show extremely complex behavior when compressed. Since Wigner predicted the dissociation and metallization of solid hydrogen under MPa pressure a century ago, predecessors have made a lot of efforts to explain many unusual characteristics of dense hydrogen, including rich and little-known solid-state polymorphism, abnormal melting line and possible superconducting properties (the application of solid-state hydrogen in superconductivity has always been a hot topic in theoretical calculation, among which the most representative ones are the superconducting properties predicted by H2S and LaH 10, and then verified by experiments). Experiments carried out under such extreme conditions are challenging, which usually lead to unexplained and controversial observation results, while theoretical research is limited by the huge calculation cost of quantum mechanics.

1. In this paper, the author uses machine learning to "learn" the potential energy surface and the interatomic force from the reference calculation, and then makes a prediction at a lower calculation cost, overcoming the limitation of length and time scale, and makes a theoretical study on the phase diagram of dense hydrogen.

2. The simulation based on machine learning potential function provides evidence for the continuous transition from molecules to atoms in liquid hydrogen, because no first-order transition is observed above the melting line. This result reveals that there can be a smooth transition between the insulating layer and the metal layer of the giant gas planet, and it can also reconcile the existing differences between experiments as evidence of supercritical behavior.

▲ Figure 1. The thermodynamic properties of hydrogen under high pressure are predicted by MLP (machine learning potential) based on PBE density functional.

Key points: a, the black curve is the predicted solid-liquid storage line, and the error line indicates the upper and lower limits of lag; The purple and orange dots represent the maximum density (ρ) and molar heat capacity (CP) at different pressures, respectively. The green dashed line and the dashed line are the * * * storage line and phase separation line of atomic and molecular fluids predicted by the polymorphic solution model, respectively; The intersection of two green curves (marked with green asterisk) is the predicted critical point of liquid-liquid transition; There are also previous experimental results in the picture. B, the purple curve represents the density isobar; The orange curve shows the molar heat capacity under different pressures; The shaded area indicates that the condition of solid phase stability corresponds to the solid-liquid storage line shown in Figure A; Error bars represent statistical uncertainty. C, at each given pressure (black line), shows the crystal structure, space group and protocell size of solid hydrogen with the lowest enthalpy.

▲ Figure 2. Fitting of polymorphic solution model and high pressure hydrogen system

Key points: A. The points in the figure represent the calculated Gibbs free energy distribution g(x) as a function of molecular fraction parameters, and the results come from eight-component kinetic simulation. Independent smoothing curve is the fitting of polymorphic solution model, which corresponds to the results obtained when T = 600, 800, 900, 1000, 1200, 1500, 1700, 2000, 2500 and 3000 K respectively. Through calculation and prediction, with the increase of temperature, the minimum value of g(x) moves to a lower molecular fraction. B. The points in the diagram represent the fitting of the solution model δ g = GM-GA. C. The points in the figure are ω values obtained by fitting g(x) to solution models at different pressures and temperatures, and curves are fitted to these values. The green dashed line corresponds to the split-phase line with ω= 2T, and the dashed line corresponds to δ g = 0, that is, the * * * storage line. The error bars in the figure are estimated from the average error of eight groups of simulations.

Original link:/articles/s 41586-020-2677-y