Deep Variational Free Energy Approach to Dense Hydrogen

Hao Xie,Zi-Hang Li,Han Wang,Linfeng Zhang,Lei Wang
DOI: https://doi.org/10.1103/PhysRevLett.131.126501
2023-09-25
Abstract:We developed a deep generative model-based variational free energy approach to the equations of state of dense hydrogen. We employ a normalizing flow network to model the proton Boltzmann distribution and a fermionic neural network to model the electron wave function at given proton positions. By jointly optimizing the two neural networks we reached a comparable variational free energy to the previous coupled electron-ion Monte Carlo calculation. The predicted equation of state of dense hydrogen under planetary conditions is denser than the findings of ab initio molecular dynamics calculation and empirical chemical model. Moreover, direct access to the entropy and free energy of dense hydrogen opens new opportunities in planetary modeling and high-pressure physics research.
Strongly Correlated Electrons,Machine Learning,Computational Physics
What problem does this paper attempt to address?
This paper aims to solve the problem of predicting the equations of state (EOS) of dense hydrogen under high pressure. Specifically, the paper proposes a variational free - energy method based on a deep generative model to study the equations of state of dense hydrogen under planetary conditions. This method jointly optimizes two neural networks - one for modeling the Boltzmann distribution of protons and the other for modeling the electron wave function given the proton positions, thereby achieving a variational free - energy level comparable to previous coupled electron - ion Monte Carlo (CEIMC) calculations. The paper points out that hydrogen is one of the most abundant elements and also the simplest one in the periodic table, consisting of only one proton and one electron. Despite its simple structure, under high - pressure conditions, dense hydrogen exhibits rich physical phenomena, such as metallization and high - temperature superconductivity. A deep understanding of these phenomena is crucial for multiple disciplinary fields such as planetary physics and nuclear fusion. Therefore, accurately predicting the equations of state and phase diagrams of dense hydrogen has always been an important test for computational methods. Traditional computational methods, such as ab initio molecular dynamics (AIMD) and machine - learning potential surfaces, although able to handle larger systems and longer time scales, their reliability still depends on the underlying electronic structure models. While the quantum Monte Carlo (QMC) method is more reliable, its high computational complexity and the problem of statistical noise limit its application range. To this end, the method proposed in the paper overcomes these problems by using deep generative models, especially normalizing flow networks and fermionic neural networks, thus providing a new and more reliable computational approach. In addition, this method directly obtains the entropy and free energy of dense hydrogen, opening up new opportunities for planetary modeling and high - pressure physics research. The paper also shows physical quantities such as free energy, energy, entropy, and pressure obtained by this method under specific conditions (\(r_s = 1.25\) and \(T = 6000 \, \text{K}\)) and compares them with the existing CEIMC results, showing good consistency. This indicates that the proposed method can not only provide more accurate physical quantity predictions but also effectively handle the complex behavior of dense hydrogen under high pressure.