Deep variational free energy prediction of dense hydrogen solid at 1200K

Xin Dong, Hao Xie, Yixiao Chen, Wenshuo Liang, Linfeng Zhang, Shenyi Liu, Han WangarXiv 2025 · DOI · arXiv

Abstract

We perform deep variational free energy calculations to investigate the dense hydrogen system at 1200 K and high pressures. In this computational framework, neural networks are used to model the free energy through the proton Boltzmann distribution and the electron wavefunction. By directly minimizing the free energy, our results reveal the emergence of a crystalline order associated with the center of mass of hydrogen molecules at approximately 180 GPa. This transition from atomic liquid to a molecular solid is marked by discontinuities in both the pressure and thermal entropy. Additionally, we discuss the broader implications and limitations of these findings in the context of recent studies of dense hydrogen under similar conditions.

Introduction

Hydrogen's behavior under extreme pressure is a cornerstone of condensed matter physics, particularly in the context of planetary interiors. At high temperatures and pressures, hydrogen is expected to transition from a molecular state to an atomic, metallic state. However, the exact nature of this transition—whether it is a liquid-liquid transition or involves a solid phase—remains a subject of intense debate. Standard Density Functional Theory (DFT) often lacks the precision to distinguish between subtle phases, while high-accuracy methods like Quantum Monte Carlo (QMC) are computationally prohibitive for extensive configuration sampling.

This study addresses these challenges by applying a deep learning-based variational free energy framework. Our findings suggest that at 1200K and approximately 180 GPa, dense hydrogen forms a molecular crystalline solid, stabilized by many-body electron correlations that are often overlooked by standard DFT functionals.

Methods

The researchers developed a deep variational free energy framework that optimizes two interconnected neural networks to minimize the total free energy ($F = E - TS$).

  1. Proton Distribution: A flow-based generative model (Normalizing Flows) learns to transform a simple base distribution into the complex Boltzmann distribution of the protons.
  2. Electron Wavefunction: A neural network ansatz (Ferminet/Deepsolid) calculates the internal energy. This architecture was enhanced with a gated attention layer to better capture long-range electron correlations.

The workflow begins with pretraining a machine-learned force field (DeePMD) using DFT data to provide initial proton configurations. This is followed by a joint variational optimization that finds the global minimum of the free energy landscape, ensuring the results are not biased by the initial DFT data.

Workflow of the deep variational free energy framework.

The model was benchmarked against high-precision data for 64 hydrogen atoms and evaluated using a simulation box of 32 atoms. The reference dataset for DeePMD included over 28,000 labeled configurations generated via concurrent learning.

Results

The variational approach revealed a spontaneous development of crystalline order at a density parameter of $r_s \approx 1.42$ (~180 GPa). This transition from an atomic liquid to a molecular solid is a first-order transition, characterized by distinct discontinuities in thermodynamic properties.

Thermodynamic signatures of the phase transition, showing pressure and entropy jumps at $r_s \approx 1.42$.

Key findings include:

  • Thermodynamic Discontinuities: Significant "jumps" in both pressure and thermal entropy at the transition point confirm the first-order nature of the phase change.
  • Structural Evidence: Radial Distribution Function (RDF) analysis showed a distinct molecular peak at 1.4 Bohr, while simulated X-ray diffraction (XRD) patterns exhibited sharp delta peaks, providing a "smoking gun" for a crystalline solid.
  • Accuracy: The neural network wavefunction achieved an accuracy of ~1 milli-Hartree per atom, comparable to gold-standard QMC methods.

Discussion and Conclusion

The discovery of a solid phase at 1200K challenges the prevailing assumption of a liquid-liquid transition in this regime. It highlights the necessity of "QMC-level" accuracy in energy functionals to capture the many-body correlations that stabilize the solid phase.

While the study provides significant insights, it is limited by the small system size (32 atoms) and a fixed cubic unit cell, which may influence the exact transition pressure and crystal symmetry. Furthermore, this work treated protons as classical particles; future research will incorporate Nuclear Quantum Effects (NQE) to provide a more complete picture of hydrogen's phase diagram. Despite these limitations, the variational framework demonstrates a powerful new path for exploring extreme states of matter with high precision.

Summary updated

Abstract

We perform deep variational free energy calculations to investigate the dense hydrogen system at 1200 K and high pressures. In this computational framework, neural networks are used to model the free energy through the proton Boltzmann distribution and the electron wavefunction. By directly minimizing the free energy, our results reveal the emergence of a crystalline order associated with the center of mass of hydrogen molecules at approximately 180 GPa. This transition from atomic liquid to a molecular solid is marked by discontinuities in both the pressure and thermal entropy. Additionally, we discuss the broader implications and limitations of these findings in the context of recent studies of dense hydrogen under similar conditions.
BibTeX
@misc{dong2025deep,
  author = {Xin Dong and Hao Xie and Yixiao Chen and Wenshuo Liang and Linfeng Zhang and Shenyi Liu and Han Wang},
  title = {{Deep variational free energy prediction of dense hydrogen solid at 1200K}},
  year = {2025},
  eprint = {2501.09590},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.09590},
  howpublished = {arXiv},
}
Figure