Machine learning assisted crystallographic reconstruction from atom probe tomographic images.

Area of Science:

• Materials Science
• Data Science
• Crystallography

Background:

• Atom probe tomography (APT) provides 3D atomic-scale compositional analysis but struggles with accurate crystallographic reconstruction.
• Recovering precise crystallographic information from APT data remains a significant challenge due to the technique's inherent limitations.

Purpose of the Study:

• To develop and validate a novel computational approach for reconstructing crystallographic information from APT data.
• To leverage deep learning for enhanced analysis of nanoscale materials using APT.

Main Methods:

• A modified forward simulation process generated a large dataset (100,000 images) of Al single crystals.
• Three deep learning models—Convolutional Neural Network (CNN), Vision Transformer (ViT), and Variational Autoencoder (VAE)—were trained for crystal reconstruction.
• The Vision Transformer (ViT) model demonstrated superior performance in recovering crystalline orientation.

Main Results:

• The ViT model achieved high accuracy in recovering crystallographic orientation, with R² values of 0.93, 0.97, and 0.93 for rotation angles φ, ψ, and θ, respectively.
• Mean Percent Errors (MPE) were as low as 0.35% for the ψ angle, indicating precise orientation recovery.
• The study validates the effectiveness of deep learning in extracting detailed crystallographic data from APT.

Conclusions:

• Deep learning models, particularly ViT, can effectively recover crystallographic information from APT data.
• This approach overcomes previous limitations in APT data analysis, enabling more accurate nanoscale material characterization.
• The findings pave the way for advanced artificial intelligence applications in atom probe tomography.