Reward Driven Workflows for Unsupervised Explainable Analysis of Phases and Ferroic Variants From Atomically Resolved...

Area of Science:

• Materials Science
• Data Science
• Physics

Background:

• Aberration-corrected electron microscopy generates complex imaging data requiring advanced analysis.
• Unsupervised machine learning (ML) methods are crucial for materials structure identification but sensitive to hyperparameter choices.
• Optimizing these ML workflows is essential for accurate materials characterization.

Purpose of the Study:

• To explore the impact of descriptors and hyperparameters on unsupervised ML for materials structure analysis.
• To develop a reward-driven approach for optimizing ML hyperparameters in electron microscopy data.
• To identify polarization and lattice distortion in Sm-doped BiFeO3 (BFO) thin films.

Main Methods:

• Applied unsupervised machine learning (ML) techniques to analyze electron microscopy imaging data.
• Developed a reward-driven optimization strategy for ML hyperparameters, focusing on domain wall continuity and straightness.
• Utilized an optimized variational autoencoder (VAE) to disentangle structural factors of variation.

Main Results:

• Demonstrated that a reward-driven approach effectively optimizes hyperparameters for materials structure analysis.
• Successfully identified local descriptors aligned with physical behavior, revealing polarization and lattice distortion in BFO thin films.
• Showcased the use of well-defined rewards as a quantifiable measure of workflow success.

Conclusions:

• A reward-driven ML workflow enhances the analysis of electron microscopy data for materials science.
• This approach provides insights into the fundamental physics of materials by aligning analysis with physical behavior.
• Optimized hyperparameter selection is critical for robust materials structure identification and characterization.