Discriminative Cooperative Networks for Detecting Phase Transitions

Area of Science:

• Physics and Computer Science
• Machine Learning Applications in Physical Sciences

Background:

• Classifying states of matter and phase transitions is a complex machine learning challenge.
• Existing algorithms often require labeled data or are not optimized for physical data analysis.

Purpose of the Study:

• To develop an unsupervised machine learning scheme for detecting phase transitions.
• To analyze new algorithms for physical data that are not typically considered in computer science.

Main Methods:

• Introduced a pair of discriminative cooperative networks (DCNs) for unsupervised phase transition detection.
• Employed an active contour model (snake) from computer vision to host the DCNs in 2D parameter spaces.
• The snake actively moves and learns within the parameter space to locate phase boundaries.

Main Results:

• The DCNs successfully detected phase transitions from fully unlabeled data.
• The scheme demonstrated efficiency in handling two-dimensional parameter spaces.
• The active contour model with DCNs automatically located phase boundaries.

Conclusions:

• The proposed unsupervised machine learning scheme offers an efficient method for identifying phase transitions.
• This approach advances the application of machine learning in analyzing physical systems and phase diagrams.
• The integration of computer vision techniques enhances the automated discovery of physical phenomena.