An improved deep learning-based algorithm for 3D reconstruction of vacuum arcs.

Area of Science:

• Plasma physics
• Computational physics
• Deep learning applications

Background:

• Deep learning methods for 3D plasma reconstruction face limitations in generalization due to insufficient training data.
• Traditional iterative methods, while robust, can be computationally intensive.

Purpose of the Study:

• To develop an improved algorithm for 3D plasma reconstruction that combines the strengths of deep learning and iterative techniques.
• To enhance the accuracy and generalization ability of plasma reconstruction, particularly for complex arc shapes.

Main Methods:

• Proposed a hybrid algorithm: convolutional neural network-maximum likelihood expectation maximization-split-Bergman (CNN-MLEM-SB).
• Utilized CNN predictions as initial values for the MLEM-SB iterative algorithm.
• Experimentally validated the method on vacuum arcs with and without transverse magnetic field (TMF) control.

Main Results:

• The CNN-MLEM-SB algorithm demonstrated superior reconstruction accuracy compared to standalone CNN methods for various arc shapes.
• Achieved high Structural Similarity Index measurement (SSIM) for simple arcs (0.952) and improved SSIM (0.868) for complex TMF-controlled arcs.
• The proposed algorithm showed increased reconstruction efficiency by 38.24% and 35.36% for disk and TMF-controlled vacuum arcs, respectively.

Conclusions:

• The CNN-MLEM-SB algorithm effectively addresses the generalization limitations of pure deep learning methods in 3D plasma reconstruction.
• This hybrid approach offers a significant improvement in both accuracy and computational efficiency for reconstructing complex plasma phenomena.