Neural Wave Functions for High-Pressure Atomic Hydrogen
David Linteau, Saverio Moroni, Giuseppe Carleo, Markus Holzmann
TL;DR
The study advances neural quantum states for electron–proton systems by integrating a backflow-enhanced orbital structure with a two-graph message-passing neural network (MPNN) that couples electrons and protons, enabling simultaneous sampling of electronic and nuclear degrees of freedom. The approach yields Born–Oppenheimer energies at up to $N=128$ atoms that match or surpass prior DMC/RMC benchmarks using VMC, and it remains accurate when incorporating nuclear quantum effects beyond the BOA by explicitly sampling proton dynamics. It preserves translational invariance without enforcing crystalline symmetry, and employs twist-averaged boundary conditions to control finite-size shell effects, enabling application to ultra-high-density hydrogen and the onset of crystal formation and pressure-induced melting. Overall, the method opens a path to accurate, symmetry-free exploration of high-pressure phase diagrams, isotope effects, and nuclear-spin phenomena in hydrogen and related systems.
Abstract
We leverage the power of neural quantum states to describe the ground state wave function of solid and liquid atomic hydrogen, including both electronic and protonic degrees of freedom. For static protons, the resulting Born-Oppenheimer energies are consistently comparable to or lower than all previous projector Monte Carlo results for systems containing up to $128$ hydrogen atoms. The same level of accuracy is preserved upon inclusion of nuclear quantum effects, thus going beyond the Born-Oppenheimer approximation. In addition, our description overcomes major limitations of current wave functions, notably by avoiding any explicit symmetry assumption on the expected quantum crystal, and sidestepping efficiency issues of imaginary time evolution with disparate mass scales. As a first application, we examine crystal formation in an extremely high-density region up to pressure-induced melting.
