Unsupervised Phase Discovery with Deep Anomaly Detection.

Area of Science:

• Condensed Matter Physics
• Quantum Many-Body Systems
• Machine Learning Applications

Background:

• Exploring complex phase diagrams is crucial for discovering new quantum phases.
• Traditional methods often require predefined labels or extensive computational resources.
• Unsupervised machine learning offers a novel approach to data-driven phase diagram exploration.

Purpose of the Study:

• To develop and demonstrate an automated, unsupervised machine learning method for exploring quantum phase diagrams.
• To identify novel or unexpected phases within the extended Bose Hubbard model.
• To utilize both theoretical (entanglement spectra) and experimental (correlation functions) data as input for machine learning.

Main Methods:

• Anomaly detection using deep neural networks on entanglement spectra and tensor network data.
• Extension of the method to utilize experimentally accessible low-order correlation functions.
• Unsupervised learning to identify deviations from normal phases within the phase diagram.

Main Results:

• Successfully mapped the entire phase diagram of the extended Bose Hubbard model in one dimension.
• Discovered a previously unidentified phase-separated region between supersolid and superfluid phases.
• Identified standard phases including superfluid, Mott insulator, Haldane-insulating, and density wave phases.

Conclusions:

• Unsupervised machine learning provides an efficient and automated pathway to explore complex quantum phase diagrams.
• The developed method can reveal unexpected phases and regions of interest.
• This approach bridges theoretical calculations and experimental data for phase discovery.