Machine learning enhanced atom probe tomography analysis

Yue Li, Ye Wei, Alaukik Saxena, Markus Kühbach, Christoph Freysoldt, Baptiste GaultProgress in Materials Science 2025 · OpenReview

TL;DR

This paper reviews the burgeoning field of machine learning (ML) applied to atom probe tomography (APT) data. It highlights how ML can automate complex analyses, overcome user bias, and unlock deeper insights into material microstructures, addressing critical challenges in standardization and data interpretation within APT.

What Problem Does This Paper Address?

Atom probe tomography (APT) is a powerful technique for 3D elemental mapping at near-atomic resolution, generating massive datasets (billions of ions). However, analyzing this data is complex, time-consuming, and heavily reliant on user expertise, leading to inconsistencies and hindering reproducibility. Current practices struggle with efficient data processing, standardization, and adherence to FAIR data principles (Findable, Accessible, Interoperable, Reusable). This paper addresses the need for automated, robust, and user-independent analysis methods for APT data, a gap that machine learning is well-positioned to fill.

Methods

The paper provides a comprehensive overview of ML algorithms applicable to APT, categorizing them into:

  • Supervised Learning: Models trained on labeled data, useful for tasks like elemental identification from mass spectra. Decision trees (DTs) are highlighted for their interpretability and speed.
  • Self-Supervised Learning: Models that learn from unlabeled data, crucial for APT's large, unannotated datasets. This includes:
  • Clustering: Grouping data points (atoms) based on similarity, used for identifying solute clusters or phases.
  • Dimensionality Reduction: Compressing data into lower dimensions while preserving information, often using autoencoders.
  • 3D Point-Cloud-Based Algorithms: Techniques specifically designed for the XYZ coordinate data of APT, including:
  • Segmentation in 3D: Dividing point clouds into meaningful subsets (e.g., identifying specific features).
  • Feature Extraction: Identifying key points, edges, or corners within the data.
  • Object Recognition: Detecting and classifying objects within the point cloud.

The review details specific ML applications in APT, such as mass spectrometry analysis, crystallographic information extraction, chemical ordering analysis, and microstructural feature detection.

Datasets

The paper discusses the application of ML to APT datasets, which are generated from experimental APT instruments. It references various studies that utilize these datasets for training and evaluation. No new datasets or benchmarks are explicitly introduced in this review paper.

Key Insights and Ideas

  • Bridging the Gap: ML offers a transformative opportunity to standardize APT data analysis, reduce user dependency, and improve reproducibility, directly addressing the limitations of manual interpretation.
  • Unlocking Hidden Patterns: ML algorithms can uncover intricate microstructural features and subtle chemical ordering that might be missed by human analysis, leading to new scientific discoveries.
  • FAIR Data Integration: ML can facilitate the adoption of FAIR data principles by enabling automated analysis and semantic representation of APT data.
  • Beyond Human Capability: ML can handle the sheer volume and complexity of APT data, enabling analyses that are computationally infeasible or too time-consuming for manual approaches.
  • Specialized Algorithms for 3D Data: The development of 3D point-cloud-based ML algorithms is crucial for effectively analyzing the inherent 3D nature of APT data.

Key Results

The paper highlights the successful application of ML in various APT analysis tasks:

  • Mass Spectrometry Analysis: Bayesian approaches and decision tree-based models (ML-ToF) have been developed for automated peak identification and elemental/molecular assignment in ToF-MS spectra, improving accuracy and efficiency.
  • Crystallographic Analysis: ML-APX, using deep neural networks, demonstrates robust crystallographic orientation extraction from detector hit maps, even in the presence of noise, outperforming traditional Hough transform methods.
  • Chemical Ordering: ML-APT, utilizing 1D/2D CNNs, has enabled quantitative analysis of chemical short-range ordering (CSRO) and medium-range ordering (CMRO) at near-atomic scales, pushing the size limit of recognized CSRO domains below 1 nm. 3D CNNs are also being explored for more comprehensive analysis.
  • Microstructural Feature Detection: ML techniques are being used for automated identification and quantification of grain boundaries, phase segmentation, and detection of complex microstructural features like dislocations, improving robustness and reducing bias.

Limitations and Future Work

  • Data Quality and Resolution: APT's inherent anisotropic spatial resolution and detection efficiency can still limit the accuracy of ML-based analyses, particularly for very small features like CSRO.
  • Computational Resources: Handling large 3D point cloud data requires significant computational power, necessitating optimization and potentially hardware acceleration (e.g., GPUs) for commercial platforms.
  • User Expertise for ML: Selecting appropriate ML model parameters can be challenging for APT users without deep ML knowledge, highlighting the need for user-friendly interfaces and guided parameter selection.
  • Standardization and Documentation: While ML offers automation, there's a need for better documentation of ML workflows and standardization to ensure reproducibility and interoperability, aligning with FAIR principles.
  • Future Directions: The authors suggest continued integration of ML into standardized workflows, improving data quality, accelerating simulations (e.g., through ML-enhanced interatomic potentials for MD simulations), and developing digital twins for APT experiments. Interfacing existing ML tools with commercial platforms is also a key area for future development.

Summary updated

BibTeX
@inproceedings{li2025machine,
  author = {Yue Li and Ye Wei and Alaukik Saxena and Markus Kühbach and Christoph Freysoldt and Baptiste Gault},
  title = {{Machine learning enhanced atom probe tomography analysis}},
  booktitle = {Progress in Materials Science},
  year = {2025},
  url = {https://openreview.net/forum?id=9h7DFcXoRX},
}
Figure