Calculating a Phase Diagram of a Simple Water Model Using Unsupervised Machine Learning on Simulation Data.

Area of Science:

• Computational physics
• Materials science
• Machine learning applications

Background:

• Phase diagrams are crucial for understanding material properties.
• Traditional methods for phase diagram determination can be labor-intensive and require significant prior knowledge.
• Developing automated and data-driven approaches is essential for complex systems.

Purpose of the Study:

• To apply unsupervised machine learning for constructing a phase diagram of a 2D rose water model.
• To evaluate the efficacy of combining dimensionality reduction and clustering algorithms for this task.
• To demonstrate a method requiring minimal prior system knowledge.

Main Methods:

• Utilized unsupervised machine learning, specifically a combination of dimensionality reduction and clustering algorithms.
• Employed two distinct datasets derived from simulations: angular distribution functions and various thermodynamic, dynamic, and structural properties.
• Compared machine learning-derived phase diagrams against manually determined ones for validation.

Main Results:

• The machine learning approach successfully predicted the phase diagram of the 2D rose water model.
• Phase diagrams generated from the two different datasets showed semiquantitative agreement.
• Identified four distinct solid phases, one liquid phase, and one gaseous phase.

Conclusions:

• The presented unsupervised machine learning method is effective and straightforward for phase diagram construction.
• This technique requires minimal prior knowledge of the system, making it broadly applicable.
• The method can also detect subtle differences within phases, aiding in the identification of local anomalies.