Ground State Phases of the Two-Dimensional Electron Gas with a Unified Variational Approach

Basic Information

• Paper ID: 2405.19397
• Title: Ground state phases of the two-dimension electron gas with a unified variational approach
• Authors: Conor Smith, Yixiao Chen, Ryan Levy, Yubo Yang, Miguel A. Morales, Shiwei Zhang
• Classification: cond-mat.str-el cs.LG physics.comp-ph quant-ph
• Publication Date: May 31, 2024 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2405.19397

Abstract

The two-dimensional electron gas (2DEG) is a fundamental model that has attracted considerable attention due to recent advances in experimental and theoretical studies of two-dimensional materials. Current understanding of the 2DEG ground state relies on quantum Monte Carlo calculations based on comparisons of different variational ansätze for different phases. This work employs a single variational ansatz—a general backflow-type wave function based on a message-passing neural quantum state architecture—to provide a unified description across the entire density range. Variational optimization consistently yields ground state energies lower than previous best results. The transition to the Wigner crystal (WC) phase occurs automatically at rs = 37±1, a lower density than currently believed. Between the liquid and WC phases, the same ansatz and variational search strongly suggest the existence of an intermediate state over a broad density range, characterized by enhanced short-range nematic spin correlations.

Research Background and Motivation

Problem Definition

This study addresses the accurate description of the ground state phase diagram of the two-dimensional electron gas (2DEG), specifically:

1. Unified Description Challenge: Existing methods require different variational ansätze for different phases (Fermi liquid FL and Wigner crystal WC)
2. Phase Transition Precision: Current estimates of the critical density for the liquid-crystal phase transition contain significant uncertainties
3. Intermediate Phase Existence: Theory predicts possible intermediate phases between FL and WC, but numerical evidence is lacking

Research Significance

• Fundamental Physics: 2DEG is a fundamental model in condensed matter physics, and its ground state energy forms the basis for density functional theory
• Experimental Relevance: Experimental advances in two-dimensional materials make 2DEG research directly applicable
• Computational Methodology: Provides new technical pathways for numerical studies of quantum many-body systems

Limitations of Existing Methods

1. Segregated Treatment: Traditional DMC methods require specially optimized trial wave functions for different phases
2. Minute Energy Differences: Energy differences between candidate states are extremely small, requiring systematically improvable and balanced accuracy
3. Limited Expressiveness: The expressiveness of traditional variational ansätze is constrained by available computational resources

Core Contributions

1. Proposed (MP)²NQS Method: Developed a unified variational ansatz based on message-passing neural networks capable of automatically discovering different phases
2. Corrected Phase Transition Critical Point: Revised the critical density for Wigner crystal transition from rs = 31±1 to rs = 37±1
3. Discovered Intermediate Phase: Found an intermediate liquid state with short-range nematic spin correlations (NSCL) in the rs ∼ 10-35 range
4. Realized "Floating" Crystal: Discovered a "floating" crystal state in the WC phase with partially restored translational symmetry
5. Exceeded Current Precision: Achieved variational energies lower than state-of-the-art DMC methods at all densities

Methodology Details

Task Definition

Study the Hamiltonian of the two-dimensional electron gas: $H = -\frac{1}{2}\sum_i \nabla_i^2 + \sum_{i<j} \frac{1}{|\mathbf{r}_i - \mathbf{r}_j|} + \text{b.g.}$

where the system is completely determined by a single dimensionless parameter rs (the Wigner-Seitz radius in units of the Bohr radius).

Model Architecture

(MP)²NQS Ansatz Structure

The wave function adopts a Slater-Jastrow-backflow form: $\Psi(\mathbf{R}) = \det[\{\langle \mathbf{x}_j;\sigma_j|\phi_a;\chi_a\rangle\}] \exp(U_2(\mathbf{R}) + U(\mathbf{R}))$

where:

• Quasiparticle Coordinates: $\mathbf{x}_j = \mathbf{r}_j + \mathcal{N}(\mathbf{R})$
• Multiplane Wave Orbitals: $\phi_a(\mathbf{x}) = \sum_{k=1}^{N_k} c_{ak} \exp(i\mathbf{G}_k \cdot \mathbf{x})$, where $N_k \gg N$

Message-Passing Network

The backflow transformation $\mathcal{N}(\mathbf{R})$ is based on a message-passing architecture:

1. Feature Definition:
   • Monadic features: $v_i = \emptyset$ (preserving translational symmetry)
   • Dyadic features: $v_{ij} = [\cos(2\pi A^{-1}\mathbf{r}_{ij}), \sin(2\pi A^{-1}\mathbf{r}_{ij}), \|\sin(\pi A^{-1}\mathbf{r}_{ij})\|, s_{ij}]$
2. Message-Passing Updates: $h_i^{(t)} = F_1^{(t)}\left(\sum_{j \neq i} m_{ij}^{(t)}, g_i^{(t)}\right) + h_i^{(t-1)}$$h_{ij}^{(t)} = F_2^{(t)}(m_{ij}^{(t)}, g_{ij}^{(t)}) + h_{ij}^{(t-1)}$
3. Attention Mechanism: $m_{ij}^{(t)} = A_{ij}^{(t)}(\{g_{ij}^{(t)}\}) \odot F_m^{(t)}(g_{ij}^{(t)})$

Technical Innovations

1. Unified Architecture Design: A single ansatz automatically adapts to liquid and crystalline phases without requiring prior specification of phase properties
2. Multiplane Wave Strategy: Design with $N_k = 3N$ allows simultaneous description of localized (crystalline) and delocalized (liquid) states
3. Enhanced Jastrow Term: Introduction of an additional Jastrow term $U(\mathbf{R})$ dependent on monadic flow and backflow coordinates
4. Optimization Stability: Employs skip connections, layer normalization, and other techniques to enhance training stability

Experimental Setup

System Parameters

• Primary System: N = 56 electrons with periodic boundary conditions
• Verification Systems: N = 58 (square cell) and N = 120 electrons
• Density Range: rs = 5-50, covering liquid to crystalline phases
• Spin Constraint: Total Sz = 0, collinear spin configurations

Optimization Methods

1. Sampling Algorithm: Metropolis-adjusted Langevin algorithm (MALA)
2. Optimizer: Stochastic reconfiguration (SR) method with improved SPRING version
3. Parallelization: 1024 Markov chains running in parallel
4. Learning Rate Schedule: Decay strategy with $\eta = \eta_0(1 + t/T)^{-1}$

Evaluation Metrics

1. Energy Accuracy: Energy comparison with DMC benchmarks
2. Structure Factor: Bragg peak identification of crystalline phases via $S(\mathbf{k})$
3. Correlation Functions: Radial distribution function $g(r)$ and spin correlations $g_s(r)$
4. Finite-Size Corrections: Extrapolation using DMC results

Experimental Results

Main Results

Energy Accuracy Improvement

• (MP)²NQS achieves lower variational energies than DMC at all densities
• Most significant improvements in the liquid phase, demonstrating the advantage of neural network ansätze in describing correlation effects
• Statistical errors smaller than line widths, demonstrating high precision of the method

Phase Transition Relocation

• Critical Density: rs = 37±1 (compared to previous rs = 31±1)
• Transition Mechanism: Precisely located through transfer learning techniques with energy crossing at rs = 36.5
• Automatic Discovery: Ansatz automatically evolves to corresponding states without prior assumptions about phase existence

Intermediate Phase Discovery

Discovery of nematic spin-correlated liquid (NSCL) in the rs ∼ 10-35 range:

• Anisotropic Spin Correlations: $g_s(\mathbf{r})$ exhibits preferred antiferromagnetic directions
• Characteristic Length Evolution: Growth from ∼1.5rs at rs = 10 to ∼1.7rs near the transition
• Charge Uniformity: Charge density remains uniform with ordering only in the spin channel

Ablation Studies

Importance of Multiplane Waves

• Low rs: Only ∼40% of k-points have significant weight (|ck| ≥ O(10^-1))
• High rs: Nearly all k-point weights are comparable, supporting crystalline state description
• Transition Region: Weight distribution evolves smoothly, demonstrating method adaptability

Backflow Effect Analysis

• Small rs: Strong position corrections, essential for liquid states
• Large rs: Weaker effects with electrons primarily maintaining initial positions
• Defect Repair: Automatically "repairs" positional defects in crystalline phases

Case Studies

"Floating" Wigner Crystal

• Partial Translational Symmetry Recovery: Density approximately uniform along antiferromagnetic stripe directions
• Theoretical Prediction Verification: First numerical observation of theoretically predicted floating crystal
• Distinction from Traditional WC: Does not prespecify electron positions in the lattice

NSCL State Characteristics

• Multiple Degenerate States: Six energetically similar states found at rs = 34 (within 0.01% range)
• Spin Density Waves: Wavelength twice that of WC
• Direction Selection: Prefers short-axis direction in rectangular cells (within ±10°)

Related Work

Traditional Quantum Monte Carlo Methods

1. Landmark Work: Ceperley & Alder (1980) established standards for electron gas research
2. 2D Systems: Systematic studies by Tanatar & Ceperley (1989), Drummond & Needs (2009)
3. Limitations: Requires specially designed trial wave functions for different phases

Neural Quantum State Development

1. Pioneering Work: Carleo & Troyer (2017) introduced the NQS concept
2. Fermionic Systems: Applications in molecular systems by Pfau et al. (2020), Hermann et al. (2020)
3. Periodic Systems: Continuous space periodic systems studied by Pescia et al. (2022)

Theoretical Predictions

1. Intermediate Phase Theory: Microemulsion phase prediction by Spivak & Kivelson (2004)
2. Experimental Observations: Intermediate states observed in MoSe₂ by Sung et al. (2023)

Conclusions and Discussion

Main Conclusions

1. The (MP)²NQS method successfully achieves unified description of 2DEG, automatically discovering different phases within a single framework
2. Wigner crystal transition occurs at lower density (rs = 37±1), revising current understanding
3. Strong numerical evidence for intermediate phase existence, supporting theoretical predictions and experimental observations
4. First numerical realization of "floating" crystal, enriching understanding of the WC phase

Limitations

1. System Size Constraints: Current calculations limited to N ≤ 120 systems, requiring larger sizes for verification
2. Finite-Size Effects: Extrapolation using DMC results may introduce systematic errors
3. Spin Constraints: Limited to collinear spin configurations, potentially missing non-collinear magnetic states
4. Optimization Landscape Complexity: Multiple local minima exist near phase transitions

Future Directions

1. Larger System Studies: Extension to larger system sizes to reduce finite-size effects
2. Non-Collinear Spins: Relaxation of collinear spin constraints to explore more complex magnetic states
3. Microemulsion Phase Research: Deeper investigation of NSCL state and its relationship to theoretical microemulsion phase predictions
4. Experimental Comparison: Quantitative comparison with experimental results from two-dimensional materials

In-Depth Evaluation

Strengths

1. Strong Method Innovation: First unified phase description with single ansatz, significant technical breakthrough
2. Deep Physical Insights: Discovered intermediate phase and floating crystal provide new physical pictures
3. High Numerical Precision: Systematically exceeds state-of-the-art methods, establishing new precision standards
4. Automatic Discovery Capability: Discovers phase transitions without human assumptions, demonstrating method universality

Weaknesses

1. Computational Cost: Neural network training computational cost may exceed traditional methods
2. Interpretability: Black-box nature of neural networks limits direct understanding of physical mechanisms
3. Parameter Sensitivity: Optimization process may be sensitive to initial conditions and hyperparameters
4. Universality Verification: Requires validation of method universality in other quantum many-body systems

Impact

1. Field Contribution: Provides new numerical tools for quantum many-body physics
2. Methodological Significance: Successful application of neural quantum states to phase transition research has demonstration value
3. Practical Value: Provides direct guidance for theoretical understanding of two-dimensional materials
4. Reproducibility: Detailed technical specifications support method reproducibility

Applicable Scenarios

1. Strongly Correlated Electron Systems: Particularly suitable for systems requiring unified description of multiple phases
2. Phase Transition Research: Unique advantages for studying quantum phase transitions
3. Ground State Property Calculations: Excellent performance when high-precision ground state energies are needed
4. Methodology Development: Provides foundation for further development of neural quantum state methods

References

1. Ceperley, D. M. & Alder, B. J. Ground State of the Electron Gas by a Stochastic Method. Phys. Rev. Lett. 45, 566–569 (1980).
2. Drummond, N. D. & Needs, R. J. Phase Diagram of the Low-Density Two-Dimensional Homogeneous Electron Gas. Phys. Rev. Lett. 102, 126402 (2009).
3. Carleo, G. & Troyer, M. Solving the quantum many-body problem with artificial neural networks. Science 355, 602–606 (2017).
4. Pescia, G. et al. Neural-network quantum states for periodic systems in continuous space. Phys. Rev. Research 4, 023138 (2022).

Note: This report is based on in-depth analysis of the full PDF text and accurately reflects the technical details and scientific contributions of the paper. This work holds significant importance at the intersection of quantum many-body physics and machine learning, providing new perspectives and tools for understanding the complex phase behavior of two-dimensional electron systems.