h1

TY  - THES
AU  - Saxena, Alaukik
TI  - Machine learning workflows for automatic analysis of atom probe tomography data
PB  - Rheinisch-Westfälische Technische Hochschule Aachen
VL  - Dissertation
CY  - Aachen
M1  - RWTH-2025-01471
SP  - 1 Online-Ressource : Illustrationen
PY  - 2024
N1  - Veröffentlicht auf dem Publikationsserver der RWTH Aachen University 2025
N1  - Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2024
AB  - The interaction between various microstructural features within a material significantly influences its macroscopic properties. Understanding and measuring the spatial aspects and composition of these microstructural features is crucial for developing new materials and investigating how their structure affects their properties. Atom probe tomography (APT) has emerged as a key technique for material characterization due to its high chemical sensitivity and ability to provide 3D mapping of atomic positions in a material at sub-nanometer resolution. However, analyzing these large datasets (often exceeding 10 million atoms) has traditionally been manual and time-consuming, leading to potential inconsistencies. The integration of machine learning, including clustering and deep learning algorithms, offers a promising solution to streamline this analysis, making it faster and less reliant on manual efforts. In response, this thesis introduces several machine learning-based workflows aimed at automatically quantifying 3D microstructures in APT datasets. These workflows strongly emphasize robustness and reproducibility, holding the potential to drive substantial advancements in the field of materials science. The initial step in analyzing the 3D microstructure within APT data involves segmenting areas with similar chemical compositions. These regions are termed chemical domains and may correspond to various phases present in a material. The segmentation process is facilitated by a novel, multi-stage unsupervised machine learning approach. To gather local composition information, the APT dataset is divided into uniformly sized voxels. The coordinate system comprising voxel compositions is called composition space. To identify distinct phases or chemical domains within this space, a Gaussian mixture model is employed for clustering. Additionally, a density-based clustering algorithm helps isolate different microstructural features within a single phase at the voxel resolution in real space, treating each feature as a distinct entity. These microstructural entities are then examined for their compositional and geometric properties (such as orientation, shape, and thickness). The methodology, along with its limitations and potential enhancements, is illustrated using both synthetic and actual APT datasets. This includes a detailed examination of a five-component, Fe-doped Sm-Co magnetic alloy and a Nickel-based superalloy that encompasses 30 distinct chemical species. Quantifying 3D microstructures can be challenging due to their intricate geometries. To address this, a data analysis pipeline using supervised machine learning is developed to examine complex microstructures containing planar subdomains, such as grain boundaries or intertwined plate-like precipitates. The pipeline allows for detailed quantification of segmented planar subdomains, both in terms of their composition (like in-plane composition fluctuations) and their geometry (such as in-plane thickness fluctuations). The robustness and efficiency of this pipeline are exemplified by its successful application to six different APT datasets corresponding to the Fe-doped Sm-Co alloy to find the correlation between material microstructure and magnetic properties. In order to segment complex linear features, particularly dislocations, a computational geometry concept called skeletonization is used to reduce the iso-composition surface meshes, delineating a microstructure into a linear graph or skeleton. The skeleton effectively encapsulates the topological information corresponding to the 3D iso-composition surface mesh. The underlying skeleton is used to segment dislocations from the meshes for detailed composition analysis. On top of this, crystallographic information from APT data is used to perform orientation analysis on each dislocation segment to transform it into the crystal coordinate system. This workflow is able to successfully extract and analyze dislocations in a Ni-based alloy and a Fe-Mn alloy, exemplifying its ability to work seamlessly on different APT datasets and its potential to be used as a powerful tool to understand the impact of decorated dislocations on material properties. Although the 3D APT data is spatially noisy, there are still signatures of the underlying crystal structure of the material in the data. Smooth overlap of atomic orbitals (SOAP) descriptors, which are translationally and rotationally invariant, were used to capture the structural and chemical information around each atom in the APT data. These descriptions are optimized using an autoencoder architecture to accommodate for the noise in the APT data. The optimized descriptors are used to train a neural network to identify ordered L1<sub>2</sub> nano-domains in an Al-Mg-Li alloy. In this thesis, a range of machine learning-based workflows and models are presented, which play a crucial role in deciphering the intricate correlations between material structure and properties. These models facilitate the extraction of descriptors related to the structure and composition of microstructures, laying the ground for training advanced, high-throughput machine learning models.
LB  - PUB:(DE-HGF)11
DO  - DOI:10.18154/RWTH-2025-01471
UR  - https://publications.rwth-aachen.de/record/1004505
ER  -