A Machine Learning Framework for Quantifying Chemical Segregation and Microstructural Features in Atom Probe...

Area of Science:

• Materials Science
• Data Science
• Computational Materials Science

Background:

• Atom probe tomography (APT) is crucial for analyzing multi-component materials.
• Quantitative analysis of APT data often requires significant human expertise for defining regions of interest.
• Understanding segregation and microstructure interplay is key in advanced materials.

Purpose of the Study:

• To introduce a computationally efficient machine learning strategy for semi-automated analysis of APT data.
• To identify compositionally distinct domains and quantify their geometric and compositional characteristics.
• To overcome limitations of manual analysis in complex material systems.

Main Methods:

• A multi-stage machine learning pipeline involving data coarse-graining into voxels and composition statistics.
• Clustering in composition space for phase identification, followed by density-based clustering for real-space segmentation.
• Refinement of segmentation using principal component analysis or U-Net-based semantic segmentation for complex morphologies.

Main Results:

• Successful semi-automated identification and segmentation of compositionally distinct phases in a Sm-(Co,Fe)-Zr-Cu alloy.
• Detailed mapping of composition distribution and segregation effects relative to precipitate geometry.
• Demonstration of disentangling intertwined, plate-like precipitate phases.

Conclusions:

• The developed machine learning approach significantly enhances the efficiency and objectivity of APT data analysis.
• This method enables detailed characterization of microstructure-composition relationships in complex materials.
• It provides a robust framework for quantitative analysis without relying solely on voxel compositions.