Sparse Mixture of Mambas for Domain Generalized Atomic Electron Tomography Augmentation

Area of Science:

• Materials Science
• Computational Imaging
• Data Science

Background:

• Atomic electron tomography (AET) is crucial for atomic-level material structure analysis.
• Raw 3D tomograms often contain artifacts due to experimental limitations like geometric constraints and low radiation doses.
• Existing artifact removal methods, while effective for simulated data, struggle with the complexity and multidomain nature of real-world tomograms.

Purpose of the Study:

• To develop a novel 3D augmentation method for enhancing domain generalization in AET.
• To address the limitations of current methods in handling complex, real-world tomograms and reduce computational demands.
• To improve the accuracy and efficiency of artifact removal in AET.

Main Methods:

• Proposed a sparse mixture of Mambas (MoMambas) framework integrating a sparse mixture-of-experts (MoE) with Mamba-based experts.
• Implemented positional information enhancement to resolve ambiguity in sparse input sequences.
• Utilized a multihead routing algorithm to optimize MoE accuracy for improved domain generalization.

Main Results:

• Achieved a 22% accuracy improvement over state-of-the-art AET augmentation methods in multidomain learning scenarios.
• Reduced the model parameter count to only 2.9% of the original.
• Lowered the computational cost by 6% compared to existing methods.

Conclusions:

• MoMambas effectively enhances domain generalization for AET artifact removal, outperforming current methods on complex datasets.
• The proposed method significantly reduces model complexity and computational overhead, making advanced AET analysis more accessible.
• The study provides a robust and efficient solution for improving the quality of 3D tomograms in materials science research.