Classifying extended, localized and critical states in quasiperiodic lattices via unsupervised learning

B. Zheng, Siyu Zhu, Xingping Zhou, Tong Liu

Abstract

Classification of quantum phases is one of the most important areas of research in condensed matter physics. In this work, we obtain the phase diagram of one-dimensional quasiperiodic models via unsupervised learning. Firstly, we choose two advanced unsupervised learning algorithms, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Ordering Points To Identify the Clustering Structure (OPTICS), to explore the distinct phases of Aubry-André-Harper model and quasiperiodic p-wave model. The unsupervised learning results match well with traditional numerical diagonalization. Finally, we compare the similarity of different algorithms and find that the highest similarity between the results of unsupervised learning algorithms and those of traditional algorithms has exceeded 98%. Our work sheds light on applications of unsupervised learning for phase classification.

Introduction

Understanding particle transport in various media is a cornerstone of condensed matter physics. While perfectly ordered crystals support extended states and completely disordered media lead to Anderson localization, quasiperiodic lattices represent a unique intermediate regime. These lattices, characterized by order without periodicity, exhibit phase transitions between extended, localized, and critical states even in one dimension. Critical states are particularly noteworthy for their multifractal nature, being neither fully spread nor fully trapped.

Traditionally, these phases are identified through metrics like the Inverse Participation Ratio (IPR) or Lyapunov exponents, which often require manual interpretation and can be ambiguous at phase boundaries. While supervised machine learning has shown promise, its reliance on labeled data limits its exploratory potential. This study investigates whether unsupervised learning can autonomously discover these quantum phases.

Methodology: Density-Based Clustering

The researchers utilized two density-based unsupervised clustering algorithms to analyze quantum eigenvectors:

DBSCAN (Density-Based Spatial Clustering of Applications with Noise)

DBSCAN groups data points based on local density. In the context of quantum states, extended states tend to form large, continuous clusters in parameter space, while localized states appear as compact, isolated clusters or "noise."

The DBSCAN clustering process: from initial uncategorized states to final classification of clusters and noise points.

OPTICS (Ordering Points To Identify the Clustering Structure)

OPTICS is an extension of DBSCAN designed to handle data with varying densities. This makes it particularly effective at identifying the subtle transitions associated with critical states, which often reside at the boundaries between extended and localized regimes.

The OPTICS algorithm's calculation sequence, identifying core points and ordering them to reveal cluster structures.

To quantify performance, a Difference Hash Algorithm—typically used in image processing—was employed to compare the AI-generated phase diagrams against traditional IPR-based benchmarks.

Models and Datasets

The algorithms were tested on two paradigmatic 1D models:

  1. Aubry-André-Harper (AAH) Model: A standard benchmark for localization transitions, using a lattice size of $L=500$.
  2. Quasiperiodic p-wave Model: A more complex system incorporating superconductivity and a broader range of critical states, using $L=1000$.

The theoretical phase boundaries of these models are well-established, providing a rigorous ground truth for validation.

Key Results and Discussion

The unsupervised approach demonstrated remarkable fidelity to traditional physics methods:

  • High Similarity: For the AAH model, DBSCAN achieved a 98.4% similarity to numerical diagonalization results. In the more complex p-wave model, similarity remained high at 95.3%.
  • Sensitivity to Criticality: While DBSCAN provided excellent broad-stroke classification, OPTICS proved superior in pinpointing the exact regions of critical states in the AAH model, which are often difficult to resolve.
  • Efficiency: The automated nature of these algorithms allows for the processing of large eigenstate datasets more efficiently than the manual calculation of Lyapunov exponents for every point in a phase diagram.

The core insight of this work is that the physical density of a wave function maps directly to the data density used by clustering algorithms. Extended states are "dense" in their connectivity, while localized states are outliers or small, tight clusters.

Conclusion

This research demonstrates that unsupervised learning can successfully categorize quantum phases without prior knowledge or labeled training data. The high degree of alignment with theoretical physics suggests these methods can be powerful tools for exploring novel materials where phase diagrams are unknown. Future work will focus on extending these techniques to many-body localized (MBL) systems, where particle interactions add significant complexity to the phase landscape.

Summary updated

Abstract

Classification of quantum phases is one of the most important areas of research in condensed matter physics. In this work, we obtain the phase diagram of one-dimensional quasiperiodic models via unsupervised learning. Firstly, we choose two advanced unsupervised learning algorithms, Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and Ordering Points To Identify the Clustering Structure (OPTICS), to explore the distinct phases of Aubry-André-Harper model and quasiperiodic p-wave model. The unsupervised learning results match well with traditional numerical diagonalization. Finally, we compare the similarity of different algorithms and find that the highest similarity between the results of unsupervised learning algorithms and those of traditional algorithms has exceeded 98\%. Our work sheds light on applications of unsupervised learning for phase classification.
BibTeX
@inproceedings{zheng2024classifying,
  author = {B. Zheng and Siyu Zhu and Xingping Zhou and Tong Liu},
  title = {{Classifying extended, localized and critical states in quasiperiodic lattices via unsupervised learning}},
  booktitle = {Chinese},
  year = {2024},
  doi = {10.1088/1674-1056/ad8cb9},
}
Figure