Predicting topological invariants and unconventional superconducting pairing from density of states and machine learning

Basic Information

• Paper ID: 2408.16499
• Title: Predicting topological invariants and unconventional superconducting pairing from density of states and machine learning
• Authors: Flávio Noronha, Askery Canabarro, Rafael Chaves, Rodrigo G. Pereira
• Classification: cond-mat.supr-con cond-mat.dis-nn cond-mat.str-el
• Publication Date: January 3, 2025
• Paper Link: https://arxiv.org/abs/2408.16499

Abstract

The competition between magnetism and superconductivity can lead to unconventional and topological superconductivity. However, experimental confirmation of the existence of Majorana edge states and unconventional pairing remains a major challenge. This paper considers a two-dimensional lattice superconductor model with spin-orbit coupling and exchange coupling to randomly distributed magnetic impurities. Depending on model parameters, the system may exhibit topologically trivial or non-trivial edge states. A phase diagram is constructed by calculating the Bott index, a topological invariant defined in real space. Machine learning algorithms are then employed to predict the Bott index from the local density of states (LDOS) at zero energy, achieving high-precision results. Machine learning models are also trained to predict the amplitude of odd-frequency pairing in the anomalous Green's function at zero energy.

Research Background and Motivation

1. Core Problem: Experimental confirmation of Majorana fermions in topological superconductors is a major challenge in current condensed matter physics. Although scanning tunneling microscopy (STM) experiments have found evidence of zero-energy states, it is difficult to distinguish between topological and trivial edge states.
2. Significance: Majorana fermions have important applications in topological quantum computing, making accurate identification of topological phases and prediction of Majorana edge state numbers crucial.
3. Limitations of Existing Methods:
   • Traditional topological invariant calculations require complete Hamiltonian information
   • It is difficult to experimentally distinguish topological from trivial edge states directly from STS data
   • Experimental detection of odd-frequency pairing is challenging
4. Research Motivation: Utilize machine learning to predict topological invariants and odd-frequency pairing from experimentally accessible LDOS data, providing new tools for experimental identification of topological superconducting phases.

Core Contributions

1. Established ML method for predicting topological invariants: First use of machine learning to predict Bott index from LDOS with accuracy exceeding 99.5%
2. Discovered effectiveness of training strategies: Demonstrated that models trained only on clean systems can predict topological phases in disordered systems
3. Predicted odd-frequency pairing: Developed linear regression models to predict odd-frequency pairing amplitude from LDOS
4. Identified intermediate topological phases: ML models can predict intermediate Bott indices B=+1 not included in the training set
5. Provided experimental application schemes: Offered practical tools for STS experiments to identify topological superconducting phases

Methodology Details

Task Definition

Input: Local density of states ρ(r,0) at zero energy, obtainable through STS experiments Output: Bott index B (topological invariant) and odd-frequency pairing amplitude F''(r,0) Constraints: Models must be trained on finite-size systems and generalize to different disorder realizations

Model Architecture

1. Physical Model

The Hamiltonian is:

H = Σ_r [ψ†_r(-tτ_z + iασ_yτ_z)ψ_{r+ê_x} + ψ†_r(-tτ_z - iασ_xτ_z)ψ_{r+ê_y} + h.c.]
    + Σ_r ψ†_r(-μτ_z + Δτ_x - J_rσ_z)ψ_r

Where:

• t: hopping parameter
• α: Rashba spin-orbit coupling
• μ: chemical potential
• Δ: superconducting pairing
• J_r: exchange coupling to magnetic impurities

2. Bott Index Calculation

Through projection operator P onto the occupied state subspace:

Û = Pe^{i2πX/l_x}P, V̂ = Pe^{i2πY/l_y}P
B = (1/2π)Im{Tr[Log(UV†U†V†)]}

3. Machine Learning Architecture

Simple Classifier (Figure 4):

• Neural network without hidden layers
• 3 output nodes corresponding to B=-1, 0, +2
• Softmax activation function
• Linear function: Z = WX + b

Complex Regressor (Figure 5):

• Two hidden layers with 600 nodes each
• ReLU activation function
• Single output node, rounded to nearest integer
• Mean squared error loss function

Technical Innovations

1. Cross-domain generalization capability: Models trained only on p=0 (clean systems) can predict topological phases in p>0 (disordered) systems
2. Weight analysis: Analysis of neural network weights reveals the model focuses on edge state features
3. Intermediate state prediction: Models can predict B=+1 states not included in the training set
4. Multi-physical quantity correlation: Established machine learning correlations between LDOS and odd-frequency pairing

Experimental Setup

Dataset

• System size: Primarily 30×30 lattices, with verification of size effects from 20×20 to 40×40
• Parameter settings: α=0.8t, Δ=1.2t, J=2t
• Disorder realizations: 47-68 different disorder realizations for each (μ,p) point
• Data partition: Training and validation sets use different disorder realizations

Evaluation Metrics

• Classification accuracy: Percentage of correctly predicted Bott indices
• Phase diagram consistency: Agreement between predicted and exactly calculated phase diagrams
• RMSE: Root mean square error for odd-frequency pairing prediction δ=0.0083

Comparison Methods

• Direct Bott index calculation as ground truth
• Comparison of different ML architectures (neural networks vs. random forests)
• Comparison of different inputs (LDOS vs. real/imaginary parts of anomalous Green's function)

Implementation Details

• Optimizer: Adam optimizer
• Loss function: Sparse categorical cross-entropy (classification) / mean squared error (regression)
• Training epochs: 5-100 epochs
• Regularization: Exclusion of transition regions near phase transitions

Experimental Results

Main Results

1. High-precision prediction: ML models achieve >99.5% accuracy in predicting Bott index outside phase transition regions
2. Cross-domain generalization: Models trained only on clean systems (p=0) can qualitatively predict phase diagram boundaries in disordered systems
3. Complete phase diagram reconstruction: Training sets including disordered systems accurately reconstruct the complete topological phase diagram
4. Intermediate state discovery: Models successfully predict B=+1 intermediate topological states despite their absence from the training set

Ablation Studies

1. Input feature comparison:
   • LDOS: Excellent prediction performance
   • Imaginary part of anomalous Green's function F''(r,0): Good performance
   • Real part of anomalous Green's function F'(r,0): Poor performance
2. Network structure impact:
   • No hidden layers: Suitable for simple classification with strong interpretability
   • Two hidden layers: Better capability for complex pattern recognition
3. Training strategies:
   • Clean systems only: Qualitatively correct but limited quantitative accuracy
   • Including disordered systems: Significantly improved prediction accuracy

Case Analysis

Weight analysis (Figure 9) shows:

• B=0 phase weights primarily concentrated in the bulk
• B=±1, ±2 phase weights primarily at edges
• Weight structure near diagonals reflects characteristics of different topological phases

Odd-frequency pairing prediction (Figure 6):

• Linear regression models accurately reproduce spatial distribution of F''(r,0)
• Edge enhancement effects are preserved in predictions
• Different topological phases exhibit different pairing patterns

Experimental Findings

1. Edge state feature identification: ML models classify topological phases primarily by recognizing LDOS enhancement at edge states
2. LDOS and odd-frequency pairing correlation: Strong linear correlation discovered, validating theoretical expectations
3. Finite-size effects: Topological phase transitions become sharper in larger systems
4. Disorder-induced phase transitions: Confirmed phase transition pathway from B=+2 to B=0 through intermediate B=+1 state

Related Work

Experimental Studies of Topological Superconductors

• Realization of magnetic atom chains on superconductor surfaces 11-13
• Two-dimensional magnetic islands and Shiba lattices 14-18
• Observation and controversy of zero-energy states in STS

Machine Learning Applications in Condensed Matter Physics

• Phase transition identification 52-63
• Topological invariant calculation 53,54,59-62
• Experimental data analysis 67-71

Odd-frequency Superconductivity

• Berezinskii's theoretical prediction 24
• Odd-frequency pairing induced by magnetic impurities 20-23
• Correlation with Majorana states 27-31

Conclusions and Discussion

Main Conclusions

1. Method effectiveness: Machine learning can predict topological invariants from LDOS with high precision, providing new tools for experimental identification of topological superconducting phases
2. Generalization capability: Models demonstrate good generalization across parameter space, particularly from clean to disordered systems
3. Physical insights: Confirmed that edge state features in LDOS are key information for distinguishing topological phases
4. Practical value: The method can be directly applied to STS experimental data analysis

Limitations

1. Model dependence: Method is based on specific theoretical models; actual materials may be more complex
2. Size constraints: Computational resources limit the system sizes that can be studied
3. Parameter range: Research focuses on specific parameter regions; generalization to other parameters requires verification
4. Experimental conditions: Noise and temperature effects in actual STS measurements are not considered

Future Directions

1. Experimental verification: Test and optimize models on real STS data
2. Model extension: Include additional physical effects such as disorder types and geometric shapes
3. Deep learning: Explore more complex neural network architectures
4. Multimodal fusion: Combine information from other experimental probes

In-depth Evaluation

Strengths

1. Strong innovation: First systematic use of ML to predict topological invariants from LDOS, opening new research directions
2. Clear physical significance: Method has solid physical foundations with clearly interpretable results
3. High experimental relevance: Directly targets experimentally accessible quantities with strong practical value
4. Complete technical approach: Detailed consideration from model construction to training strategies
5. Reliable results: Multiple verification methods ensure result reliability

Weaknesses

1. Model complexity: Possible over-engineering for relatively simple classification tasks
2. Insufficient generalization verification: While claiming applicability to real materials, corresponding verification is lacking
3. Limited physical explanation: Physical mechanisms explaining why ML models work effectively need deeper exploration
4. Computational cost: Computational demands for large-scale systems may limit application scope

Impact

1. Academic contribution: Provides new paradigm for machine learning research on topological matter
2. Practical value: May significantly improve experimental identification of topological superconducting phases
3. Methodological significance: Demonstrates effective combination of theoretical computation and machine learning
4. Field advancement: May promote more ML applications in quantum materials research

Applicable Scenarios

1. STS experimental data analysis: Direct application to scanning tunneling microscopy experiments
2. Topological material screening: Rapid identification of candidate topological superconductors
3. Theoretical computation acceleration: Replacement for time-consuming exact topological invariant calculations
4. Educational demonstration: Typical case of machine learning applications in physics

References

This paper cites 91 important references covering multiple fields including topological superconductivity, machine learning, and odd-frequency pairing, providing solid theoretical foundations for the research.

Overall Assessment: This is a high-quality theoretical physics paper that successfully applies machine learning methods to topological superconductor research. The paper's innovation, technical completeness, and practical value are all outstanding, providing valuable new tools for experimental identification of topological superconducting phases. While further verification on actual materials is needed, it has already demonstrated excellent application prospects.