U-net based vortex detection in Bose-Einstein condensates with automatic correction for manually mislabeled data.

Area of Science:

• Condensed Matter Physics
• Quantum Mechanics
• Machine Learning Applications

Background:

• Accurate localization of quantum vortices and their cores in Bose-Einstein condensates (BECs) is crucial for studying their properties.
• Traditional machine learning for vortex localization faces challenges due to the need for extensive manual data annotation and potential inaccuracies.

Purpose of the Study:

• To develop and validate an Automatic Correction Labeling (ACL) approach to improve the accuracy of vortex localization in BECs.
• To address the limitations of manual data annotation in machine learning models for quantum vortex detection.

Main Methods:

• Utilized the U-Net deep learning model for pixel-level vortex position detection.
• Implemented an Automatic Correction Labeling (ACL) process involving rough labeling, machine learning, probability region search, and data relabeling for model retraining.

Main Results:

• Demonstrated that U-Net can accurately locate vortex positions even with biased or incomplete manual annotations.
• Showcased that training U-Net with a limited dataset yields excellent predictive results due to inherent vortex regularities.
• Validated that the ACL method significantly improves machine learning model performance metrics for vortex localization by correcting labeling errors.

Conclusions:

• The ACL method effectively enhances the precision of vortex localization in BECs, particularly for vortex core identification.
• This machine learning approach offers a robust solution for accurate vortex dynamics localization and provides a framework for improving image recognition tasks with noisy labels.