Reward driven workflows for unsupervised explainable analysis of phases and ferroic variants from atomically resolved imaging data (2411.12612v1)

Abstract: Rapid progress in aberration corrected electron microscopy necessitates development of robust methods for the identification of phases, ferroic variants, and other pertinent aspects of materials structure from imaging data. While unsupervised methods for clustering and classification are widely used for these tasks, their performance can be sensitive to hyperparameter selection in the analysis workflow. In this study, we explore the effects of descriptors and hyperparameters on the capability of unsupervised ML methods to distill local structural information, exemplified by discovery of polarization and lattice distortion in Sm doped BiFeO3 (BFO) thin films. We demonstrate that a reward-driven approach can be used to optimize these key hyperparameters across the full workflow, where rewards were designed to reflect domain wall continuity and straightness, ensuring that the analysis aligns with the material's physical behavior. This approach allows us to discover local descriptors that are best aligned with the specific physical behavior, providing insight into the fundamental physics of materials. We further extend the reward driven workflows to disentangle structural factors of variation via optimized variational autoencoder (VAE). Finally, the importance of well-defined rewards was explored as a quantifiable measure of success of the workflow.

• The paper introduces a reward-based strategy that aligns unsupervised workflow hyperparameters with physical material properties in STEM imaging.
• It employs Gaussian Mixture Model clustering and variational autoencoders to accurately differentiate domain regions and structural variations in Sm-doped BiFeO₃ thin films.
• The approach streamlines real-time, explainable data analysis, reducing the need for manual parameter tuning in high-resolution microscopy.

Analysis of Reward-Driven Workflows for Unsupervised Explainable Atomically Resolved Imaging

The paper presents a method development to refine unsupervised analysis workflows for atomically resolved imaging data using a reward-driven strategy. Focused on scanning transmission electron microscopy (STEM) datasets, the paper addresses the challenge of identifying phases and ferroic variants. The approach is demonstrated particularly on Sm-doped BiFeO₃ (BFO) thin films with varying dopant concentrations to highlight its effectiveness.

The authors elucidate how their reward-driven approach offers a robust alternative to the hyperparameter sensitivity typically observed in unsupervised ML techniques such as clustering and dimensionality reduction. Specifically, a reward-driven optimization controls key hyperparameters in the analysis workflow, aligning the unsupervised ML process with physical attributes of the material.

Key Methodological Insights

1. Reward-Driven Optimization:
   • The analysis pipeline is framed as an optimization problem where reward functions are constructed to reflect domain-specific knowledge, such as domain wall continuity and straightness. This ensures alignment with physical behaviors and properties of the material, providing an insightful perspective into fundamental physics.
2. GMM-Based Clustering:
   • Gaussian Mixture Model (GMM) clustering was employed to identify domain regions from atomically centered descriptors. Window size and number of GMM components are meticulously optimized to balance over-segmentation against under-segmentation, yielding correlation with known polarization maps.
3. Variational Autoencoder (VAE) Applications:
   • The reward-driven framework is extended to the disentanglement of material structure variations using a VAE. Conditional VAE models help in isolating distinct structural features attributable to specific atomic sites, providing a comprehensive look into lattice aberrations and phase transitions.

Resulting Implications

The introduced reward concept substantiates an accessible path for real-time, explainable data analysis, which is pivotal for autonomous operations in microscopy. This algorithmic breakthrough supports the implementation of physical heuristics into ML workflows, enhancing the interpretability and relevance of condensed matter studies.

By grounding ML workflows in physical reality through reward functions, the research moves towards automating traditionally subjective aspects of high-resolution imaging analytics. This method has shown strong performance against physics-derived ground truth polarization distributions, indicating its potential to supplant manual parameter tuning.

Future Directions

This work opens avenues for further generalization of reward-driven workflows beyond microscopy. The authors suggest potential applications for the method in diverse areas of physics discovery, as long as pertinent physics-oriented reward functions can be conceived. Additional work could be directed at expanding the parameter space explored by such workflows and integrating multi-modal datasets for comprehensive material analysis.

Given the ongoing advances in hardware, combining reward-driven workflows with edge computing could mark a significant shift in how large volumes of scientific data are processed and understood in real-time, making its implications far-reaching in both the theoretical groundwork and practical applications of ML in material sciences.