Testing Outlier Detection Algorithms for Identifying Early Stage Solute Clusters in Atom Probe Tomography.

Area of Science:

• Materials Science
• Data Analysis
• Nanotechnology

Background:

• Atom probe tomography (APT) is crucial for analyzing solute clustering and precipitation in materials.
• Standard clustering algorithms like DBSCAN struggle with accurately identifying small clusters due to size limitations (Nmin parameter).

Purpose of the Study:

• To address the limitations of existing methods for cluster characterization in APT data.
• To explore outlier detection algorithms as an alternative approach for identifying solute clusters.
• To develop an improved method for cluster identification in APT datasets.

Main Methods:

• Characterizing clustering in APT data as an outlier detection problem.
• Testing k-nearest neighbors local outlier factor (k-NN LOF) and learnable unified neighborhood-based anomaly ranking algorithms.
• Comparing these methods against standard DBSCAN using simulated APT data.
• Auto-thresholding algorithm decision scores using the Karcher mean to minimize human bias.

Main Results:

• Tested outlier detection algorithms outperformed DBSCAN for clusters smaller than 25 atoms.
• These algorithms underperformed compared to DBSCAN for clusters larger than 30 atoms in simulated data.
• A novel combined k-NN and DBSCAN method demonstrated robust performance across all tested cluster sizes.

Conclusions:

• Outlier detection offers a promising alternative for APT cluster analysis, particularly for small clusters.
• The developed combined k-NN and DBSCAN approach provides a more reliable method for identifying clusters in APT data.
• This new approach enhances the characterization of solute clustering and precipitation in materials using APT.