Machine Learning is Unlocking New Potential in Atom Probe Tomography

Machine Learning is Unlocking New Potential in Atom Probe Tomography

Atom Probe Tomography (APT) is a crucial technique in materials science, providing near-atomic-level 3D reconstructions and precise chemical analysis. However, interpreting complex datasets, especially for local chemical ordering (LCO) and subtle defects, is time-consuming and requires expert-level analysis. Integrating machine learning (ML) into APT workflows significantly improves resolution, time to knowledge, and analytical power, advancing APT's performance and uncovering new microstructural insights.

Baptiste Gault, Ye Wei, Yue Li, Alaukik Saxena

Much of these advancements are being driven by Baptiste Gault and his team at the Max-Planck Institute for Sustainable Materials , Düsseldorf, Germany, a premier materials science and engineering research institute.

Gault, a permanent W2 Group Leader at Max Planck Institute for Sustainable Materials, and his team — Ye Wei , Postdoctoral Researcher, EPFL (école polytechnique fédérale de Lausanne); Yue Li , Researcher, Max-Planck-Institut für Eisenforschung GmbH; and Alaukik Saxena , Scientist at Max Planck Institute for Sustainable Materials and Fellow at Helmholtz School for Data Science — are at the forefront of combining ML techniques with APT to push the boundaries of what this technology can achieve.

Machine Learning and Atom Probe Tomography: A Synergistic Advancement

With unmatched chemical sensitivity at sub-nanometer resolution, APT is essential for examining nanoscale materials and their properties. However, compositional segregation alone may be limited in addressing subtle variations and complex defects like stacking faults and LCOs. Machine learning algorithms are overcoming these challenges, enabling the extraction of finer details from APT data.

“Part of the community has long focused on developing methodologies for ‘mining’ the 3D point cloud, such as quantifying precipitate distribution using nearest-neighbor distances and clustering, now seen as part of machine learning techniques,” states Professor Gault.

“Since 2018, my group at the Max-Planck Institute for Sustainable Materials has been developing ML-based methods to process data from atom probe tomography and field-ion microscopy. Our goal is to use physics-informed models to automate critical tasks in atom probe data reconstruction and analysis, such as identifying poles in ion detector hit maps, elemental identification in mass spectra, and segmenting microstructural features.

“Recently, we advanced the analysis of short-range order by APT, a long-standing community interest. Machine learning has proven effective in identifying specific patterns, such as Yue Li’s ML approach to quantify short-range order in binary Fe-Al and Fe-Ga, revealing 3D distributions of short-range ordered zones. This method has also been adapted for more complex alloys, revealing short-range order in a ternary medium-entropy alloy.

"We are now working on generalizing this search across entire datasets,” explains Gault.

Overcoming Compositional Challenges with ML

One challenge in APT is analyzing microstructures with minimal compositional segregation. Traditionally, compositional analysis has been used to identify precipitates and structural defects, but this approach is limited when addressing phenomena like LCOs. The 3D deep learning approach AtomNet enhances APT by using point cloud data to simultaneously identify compositional and structural information (Yu et al., 2024). This method marks a significant leap in detecting features like L12-type nanoprecipitates and L10-type LCOs, which are difficult to resolve with traditional techniques. AtomNet's ability to segment features like stacking faults in a cobalt-based superalloy without training data for the specific fault type demonstrates ML's potential to uncover hidden microstructures (Yu et al., 2024).

Enhancing Resolution Beyond Conventional Limits

Machine learning enhances segmentation and pushes APT's resolution boundaries. Li et al. (2023) introduced a machine learning-enhanced APT approach that enabled the 3D imaging of chemical short-range orders (CSROs), a feature typically beyond conventional APT analysis. This method uses convolutional neural networks (CNNs) to process spatial distribution maps (SDMs) from APT data, surpassing typical resolution limits. By applying ML, Li et al. (2023) discerned B2-CSROs in body-centered cubic Fe-Al alloys, which eluded traditional techniques.

Automated and Unbiased Analysis

ML replaces manual analysis with algorithms, uncovering complex precipitate structures often overlooked. A machine learning framework proposed by Saxena et al. (2023) integrates a multi-stage ML strategy to automate the segmentation of chemically distinct domains from APT datasets. This automation accelerates the analysis process and provides more consistent and reproducible results (Saxena et al., 2023).

Improving Quantitative Correlations and Material Design

Integrating ML into APT improves resolution and accuracy, enhancing correlations between microstructural features and material properties. For instance, the ML-enhanced approach by Li et al. (2023) allowed researchers to quantitatively correlate annealing temperature, CSRO, and mechanical properties like nano-hardness and electrical resistivity in Fe-Al alloys. These correlations inform high-performance material design. ML's ability to extract meaningful correlations from complex datasets makes it a critical tool for advancing material science research and innovation (Li et al., 2023).

Future Potential of Machine Learning in APT

While ML has proven its value in advancing APT, significant potential remains for further enhancements. Future developments may include integrating ML with other imaging techniques, such as electron microscopy, to create more comprehensive multi-modal analysis platforms. ML could better handle noisy or incomplete data as it evolves, making APT more robust in suboptimal conditions.

Another promising direction is using ML to predict material properties based on APT datasets, further streamlining materials discovery. Current research explores active learning approaches, where ML models guide data acquisition in real-time (Bauer et al., 2024), focusing on areas of greatest interest and reducing the need for extensive manual experimentation — thereby creating opportunities to explore novel processes.

Summary of typical APT data analysis workflow, from the processing of the experimental data to form and analyse the mass spectrum to the reconstruction of the 3D point cloud, all of these steps typically require user input, highlighting potential for ML-learning and possible complete data processing workflows. (Bauer et al., 2024)

Transforming Atom Probe Tomography

Machine learning transforms Atom Probe Tomography from a powerful but labor-intensive technique into a more potent, automated analytical tool. ML unlocks new potential by overcoming compositional analysis limitations, enhancing resolution, and enabling more accurate microstructure-property correlations. As ML technologies evolve, APT's capabilities will continue to advance, driving innovation in material design and discovery. The synergy between APT and ML is crucial in advancing nanoscale materials research.


CAMECA and APT Innovation

CAMECA's latest model 3D Atom Probe Microscope is the

CAMECA is a world leader in delivering breakthrough Atomic Probe Tomography technology for research and industry. Current CAMECA APT solutions include the 6000 family — the Invizo? 6000 , the LEAP 6000 XR? — and the EIKOS-UV? .

CAMECA is excited about the potential of machine learning to push the boundaries of atom probe data analysis. Active collaborations are underway with several partners, including the Max-Planck Institute for Sustainable Materials, to incorporate tools like those described here as extensions for IVAS (Integrated Visualization & Analysis Software) for CAMECA Atom Probes and APSuite 6 (collaborative platform to seamlessly manage your entire Atom Probe Tomography research projects within one single environment). By making them widely available on the ubiquitous analysis platform, the entire community can evaluate them and learn how their applications can benefit from these advanced analyses.

For additional resources and information, visit Atom Probe Tomography (APT) on the CAMECA website.?


About Professor Baptiste Gault

Professor Baptiste Gault is a permanent W2 Group Leader at Max Planck Institute for Sustainable Materials , a premier materials science and engineering research institute. He is also a part-time Professor in the Department of Materials at Imperial College, London, and an Honorary Professor in the Department of Physics, University Duisburg-Essen, Germany. His expertise is in atom probe tomography and its use to characterize advanced materials. His leadership in the field was recognized by the Leibniz Prize in 2020, the highest scientific prize in Germany.

Photo credit: Imperial College.


References

  • Yu, J. et al. (2024). 3D deep learning for enhanced atom probe tomography analysis of nanoscale microstructures. DOI: 10.1016/j.actamat.2024.120280 .
  • Bauer, S. et al. (2024). Modelling and Simulation in Materials Science and Engineering. DOI: 10.1088/1361-651X/ad4d0d .
  • Li, Y. et al. (2023). Quantitative three-dimensional imaging of chemical short-range order via machine learning enhanced atom probe tomography. DOI: 10.1038/s41467-023-43314-y .
  • Zhou, X. et al. (2022). Revealing in-plane grain boundary composition features through machine learning from atom probe tomography data. DOI: 10.1016/j.actamat.2022.117633 .
  • Saxena, A. et al. (2023). A machine learning framework to discover patterns in atom probe tomography data. DOI: 10.1093/micmic/ozad086 .

Pierre-Yves Corre

Sales & Service Manager | CAMECA (AMETEK) | Australia & Southeast Asia | Expert in Material Analysis Solutions

1 个月

Fascinating work combining Machine Learning techniques with Atom Probe! A good read Lionel Gueguen.

APT datasets are so rich in information that humans could really use a hand to wring every bit of information out of it!

要查看或添加评论,请登录

社区洞察

其他会员也浏览了