Deep learning for Gaussian process soft x-ray tomography model selection in the ASDEX Upgrade tokamak.

Area of Science:

• Plasma physics
• Machine learning
• Computational science

Background:

• Gaussian process tomography (GPT) enables real-time plasma emissivity profile reconstruction in tokamaks.
• Bayesian formalism within GPT facilitates model selection by comparing model evidence.
• High-dimensional data in model selection can lead to slow computations, hindering real-time applications.

Purpose of the Study:

• To accelerate GPT model selection for tokamak plasma diagnostics.
• To compare the efficiency of a convolutional neural network (CNN) approach against traditional Bayesian methods.
• To utilize CNN-based classifications for improved tomographic reconstructions.

Main Methods:

• Training a convolutional neural network (CNN) to map Soft X-Ray (SXR) tomographic projections to the highest-evidence GPT model.
• Utilizing ASDEX Upgrade tokamak SXR diagnostic data for network training and validation.
• Comparing the speed and accuracy of the CNN approach with analytical Bayesian evidence calculations.

Main Results:

• The CNN significantly reduces the time required for GPT model selection compared to analytical Bayesian methods.
• The CNN approach demonstrates comparable or improved accuracy in identifying the best GPT model.
• CNN-based classifications were successfully used to generate tomographic reconstructions of plasma emissivity profiles.

Conclusions:

• Convolutional neural networks offer a computationally efficient and effective alternative for Gaussian process tomography model selection in tokamak research.
• This AI-driven method enhances the feasibility of real-time plasma diagnostics and analysis.
• The trained network can be applied to SXR data for rapid model selection and subsequent plasma profile reconstruction.