First-Principles AI finds crystallization of fractional quantum Hall liquids

Technical Summary: First-Principles AI Finds Crystallization of Fractional Quantum Hall Liquids

Problem Statement
The central challenge addressed is determining the conditions under which a fractional quantum Hall (FQH) liquid crystallizes. This question lies at the heart of the competition between topological order and charge ordering in two-dimensional electron systems (2DEG) under strong magnetic fields. Existing theoretical and numerical approaches face significant limitations:

• Trial Wavefunctions: Traditional methods rely on separate, tailor-made trial wavefunctions for FQH liquids and Wigner crystals, making it difficult to unbiasedly determine phase boundaries.
• Quantum Monte Carlo (QMC): Suffers from severe complex phase problems in interacting systems.
• Density Matrix Renormalization Group (DMRG): Suffers from discretization errors due to the necessary truncation of Landau levels.
• Landau Level Mixing: In regimes with strong Landau-level (LL) mixing (controlled by the parameter $\kappa = U/K$), the ground state is governed by a competition between fractionalization and crystallization that is genuinely non-perturbative.

Methodology: MagNet
The authors introduce MagNet, a self-attention neural-network variational wavefunction designed for quantum systems in magnetic fields on a torus geometry. Key methodological innovations include:

1. Geometry: The study utilizes a torus geometry defined by basis vectors $\mathbf{L}_1$ and $\mathbf{L}_2$ pierced by a quantized magnetic flux. This choice avoids boundary effects (unlike the disk) and curvature issues (unlike the sphere), while naturally accommodating both uniform topological fluids and periodic crystalline orders.
2. Wavefunction Architecture:
   • The ansatz is constructed as a sum of determinants of generalized many-body orbitals: $\Psi = e^J \sum_n \det[\phi^n_j]$.
   • Boundary Conditions: The network explicitly enforces magnetic translation boundary conditions. The orbitals $\phi^n_j$ are factorized into a periodic amplitude $F^n_j$ and a phase factor $\chi^n_j$.
   • Winding Maps: The core innovation is the parametrization of $\chi^n_j$ via a product of functions $f(z_i - \eta)$, where $\eta$ represents "winding maps." These maps are learnable, periodic, symmetric many-body functions that determine the positions of zeros (vortices) in the wavefunction.
   • Expressivity: By allowing the zeros of the orbitals to depend on all particle coordinates (including the particle itself), the ansatz becomes non-holomorphic. This enables the representation of states outside the Lowest Landau Level (LLL) and captures non-trivial phase effects without external physics priors.
3. Training: The network is trained solely by minimizing the variational energy of the microscopic Hamiltonian. No external training data, knowledge of Landau levels, Laughlin states, flux attachment, or crystalline order is provided.

Key Results
The authors applied MagNet to the $\nu = 1/3$ filling factor across a broad range of Landau-level mixing parameters ($\kappa$):

• Weak Mixing ($\kappa = 3.0$): The network spontaneously discovers a fractional quantum Hall liquid. The pair correlation function $g(r)$ and structure factor $S(q)$ show liquid-like behavior. Crucially, the optimized wavefunction exhibits the correct topological ground state degeneracy, carrying weight only in the three center-of-mass momentum sectors expected for the $\nu=1/3$ Laughlin state.
• Strong Mixing ($\kappa \geq 20$): As $\kappa$ increases, the system transitions to a crystalline phase. The pair correlation function develops distinct spatial modulations, and the structure factor peak $S(K)$ grows linearly with particle number $N$, indicating true long-range crystalline order.
• Intermediate Regime: Finite-size scaling of the structure factor suggests that incipient charge order begins to appear around $\kappa = 12$ and $15$, with robust long-range order emerging between $\kappa = 15$ and $\kappa = 20$.
• Energy Comparison: For $\kappa = 3.0$, the neural network achieves lower variational energies than exact diagonalization (ED) projected onto the LLL. This demonstrates the advantage of the real-space NN ansatz, which naturally incorporates LL mixing without truncation, whereas LLL-projected ED is restricted to the liquid phase regardless of $\kappa$.

Significance and Claims
The paper claims to realize a "genuinely first-principles AI solver" for this paradigmatic strongly correlated problem. Its significance is framed around:

1. Unifying Framework: MagNet provides a single, unified architecture capable of describing both topological FQH liquids and electron crystals. The network discovers the correct ground state phase directly from the Hamiltonian without bias.
2. First-Principles Discovery: The method finds competing phases (liquid vs. crystal) without external training data or pre-existing physical knowledge of the phases. All information regarding fractionalization and crystallization is extracted a posteriori from the learned wavefunction.
3. Resolving Phase Competition: The work offers a unified solution to the FQH-to-crystal topological quantum phase transition, allowing for the tracking of correlation functions and the location of the onset of crystalline order in a regime where previous methods struggled.
4. General Applicability: The authors suggest this approach highlights the power of first-principles AI as a general tool for quantum matter, capable of exploring phase diagrams and discovering unexpected correlated phases in strongly interacting systems.

The paper concludes by noting that while evidence for pronounced crystalline correlations is found, the full nature of the crystal phase (e.g., whether it is a Wigner crystal, composite fermion crystal, or a Hall crystal) remains to be fully established. Future work is proposed to apply this framework to a broader set of filling factors and systems with spatially non-uniform magnetic fields, such as moiré systems.