Detecting Clusters in Atom Probe Data with Gaussian Mixture Models.

Area of Science:

• Materials Science
• Data Analysis
• Computational Methods

Background:

• Atom probe tomography (APT) reconstructions are crucial for materials analysis.
• Identifying clusters in APT data is challenging and often relies on user-defined parameters.
• Existing methods like maximum separation lack probabilistic interpretation.

Purpose of the Study:

• To develop a novel, automated algorithm for accurate cluster identification in APT data.
• To introduce a machine learning approach that probabilistically distinguishes clusters from noise.
• To provide a more robust and accurate alternative to the maximum separation method.

Main Methods:

• Developed the Gaussian mixture model Expectation Maximization Algorithm (GEMA).
• Utilized expectation maximization to learn cluster properties (position, size, width).
• Employed probabilistic atom assignment to clusters for uncertainty quantification.

Main Results:

• GEMA accurately identifies clusters by modeling them with Gaussian distributions.
• Probabilistic assignment provides meaningful uncertainty at precipitate/matrix interfaces.
• GEMA demonstrated superior cluster detection accuracy compared to the maximum separation method on simulated data.

Conclusions:

• GEMA offers a significant advancement in automated cluster analysis for APT.
• The algorithm's probabilistic nature enhances the scientific interpretation of results.
• GEMA is effective for both simulated and real-world APT datasets.