Comparison of Tomographic Reconstruction Algorithms for Infrared Imaging Video Bolometer Diagnostic in Plasma Devices

This paper evaluates and compares the performance of Minimum Fisher Information, Phillips-Tikhonov regularization, and Maximum-Likelihood Expectation-Maximization algorithms for reconstructing 2D plasma radiation emissivity from Infrared Imaging Video Bolometer data, analyzing their trade-offs in accuracy, stability, and suitability for real-time or offline applications across various viewing geometries and emissivity profiles.

The Gist

Imagine you are trying to figure out what a mysterious, glowing cloud looks like inside a dark room, but you can't see the cloud directly. All you have is a piece of paper with a tiny hole in it, placed between you and the cloud. The cloud emits light (radiation) that passes through the hole and hits the paper, leaving a blurry, smeared shadow. Your job is to look at that shadow and mathematically "reverse-engineer" the original shape and brightness of the cloud.

This is exactly what scientists do with plasma (the super-hot, glowing gas inside nuclear fusion reactors). They use a device called an Infrared Imaging Video Bolometer (IRVB). Think of the IRVB as a high-tech camera that doesn't take a picture of the plasma directly. Instead, it looks at a thin metal foil that gets heated up by the plasma's radiation. The camera measures how hot different spots on the foil get, creating a "shadow" of the plasma's heat.

The problem is that this shadow is a messy mix of all the light coming from every angle. To see the actual 3D shape of the plasma's heat, scientists have to solve a difficult math puzzle called tomography (the same math used in CT scans for the human body).

The Four "Detectives"

The paper tests four different mathematical "detectives" (algorithms) to see which one is best at solving this puzzle. The researchers created five different "fake plasma" scenarios (called phantoms) to test them, ranging from a simple glowing ball of light to complex, hollow rings and split shapes near the edges of the reactor.

Here is how the four detectives performed:

1. The "Smooth Operator" (PTR-2):

   • How it works: This method assumes the plasma is generally smooth and tries to avoid wild, jagged jumps in brightness. It's like smoothing out a crumpled piece of paper.
   • The Verdict: It is the fastest and most reliable for real-time use. It solves the puzzle in less than a second. While it's not perfect at finding tiny, sharp details, it's good enough to give a clear picture quickly. If you need to know what's happening right now in the reactor, this is your best bet.

2. The "Adaptive Specialist" (MFI):

   • How it works: This detective is smarter about where to look. It knows that some parts of the plasma are very bright and others are dim, so it adjusts its focus accordingly. It's like a photographer who automatically changes the focus depending on whether the subject is in the shadows or the sunlight.
   • The Verdict: It is the most accurate at reconstructing the true shape, especially for tricky, complex shapes like the "double-null" (a split shape) or asymmetric blobs. However, it is slow. It takes about 3 seconds to solve the puzzle. This is too slow for real-time control but perfect for detailed analysis after the experiment is over.

3. The "Basic Smoother" (PTR-1):

   • How it works: Similar to the Smooth Operator, but it uses a simpler, less flexible rule for smoothing.
   • The Verdict: It works okay for simple, round shapes but fails miserably when the plasma has complex, split, or edge-heavy shapes. It tends to blur out important details. The paper suggests skipping this one for difficult cases.

4. The "Statistical Gambler" (MLEM):

   • How it works: This method uses a specific statistical approach that assumes the light comes in "packets" (photons). It builds the image step-by-step, getting closer with every guess.
   • The Verdict: It is incredibly fast (the fastest of all), but it is unreliable. It often creates a picture that looks nothing like the real plasma, especially when the heat is concentrated at the edges. It's like a gambler who wins quickly but often loses the big prize. The paper advises against using it for this specific type of plasma camera unless the noise conditions are very specific.

The "Resolution" Trade-off

The paper also tested how the size of the puzzle pieces affects the result.

• Too few pieces (Low resolution): The picture is blurry, but you can solve it easily.
• Too many pieces (High resolution): The picture could be sharp, but you don't have enough data to fill in all the tiny gaps. The math gets confused, and the image becomes noisy or wrong.
• The Sweet Spot: The researchers found that for their specific camera setup (a 9x9 grid of sensors), a 25x25 grid for the final image is the perfect balance. Going finer than that doesn't help because the camera doesn't have enough "eyes" to see that much detail.

The Bottom Line

If you are running a nuclear fusion experiment and need to see the plasma's heat map instantly to keep the reactor safe, use the PTR-2 method. It's fast and good enough.

If you want to study the data later to understand exactly how the plasma behaved in a complex event, use the MFI method. It takes a few seconds longer, but it gives you the most accurate, high-definition picture of what actually happened.

The paper concludes that there is no single "perfect" method; it depends on whether you value speed (for real-time safety) or precision (for deep scientific analysis).

Technical Summary

Technical Summary: Comparison of Tomographic Reconstruction Algorithms for IRVB Diagnostics

Problem Statement
The Infrared Imaging Video Bolometer (IRVB) is a diagnostic tool used in magnetically confined plasma devices to measure total radiation power loss in two dimensions. Unlike discrete chord-based bolometers, the IRVB utilizes a pinhole camera geometry where a free-standing metal foil acts as a broadband absorber, imaged by an infrared camera. This setup generates line-integrated signals that must be mathematically inverted to recover the local plasma emissivity distribution on a poloidal cross-section. This inversion is an ill-posed problem, particularly challenging when dealing with sparse data (e.g., the 9×9 pixel array available on the ADITYA tokamak), limited viewing angles, and the need to preserve sharp features such as divertor strike points or asymmetric impurity accumulation. While various tomographic methods exist, there is a lack of comparative assessment specifically tailored to the unique geometry and data constraints of IRVB systems on medium-sized tokamaks.

Methodology
The study establishes a comprehensive forward-modelling and inverse-reconstruction framework to evaluate four specific tomographic algorithms:

1. First-Order Phillips-Tikhonov Regularization (PTR-1): Minimizes a penalized least-squares functional with a first-order gradient penalty.
2. Second-Order Phillips-Tikhonov Regularization (PTR-2): Similar to PTR-1 but employs a Laplacian (second-order) penalty to enforce higher-order smoothness.
3. Minimum Fisher Information (MFI): A spatially adaptive method that weights regularization inversely proportional to the emissivity estimate, concentrating smoothness in low-emission regions.
4. Maximum-Likelihood Expectation-Maximization (MLEM): An iterative algorithm maximizing Poisson log-likelihood, inherently enforcing non-negativity through multiplicative updates.

The evaluation utilizes a geometry model derived from the ADITYA-Upgrade (ADITYA-U) tokamak, discretizing the plasma into 3D voxels and projecting them onto a 2D poloidal plane under the assumption of toroidal symmetry. To test robustness, five synthetic "phantom" emissivity profiles were constructed to represent typical plasma scenarios:

• Core Gaussian (centrally peaked).
• Hollow Gaussian (annular ring).
• Low-Field-Side Asymmetric (outboard impurity accumulation).
• Double-Null Divertor (two compact blobs at strike points).
• Physical Impurity Emission (based on synthetic electron temperature and density profiles with Carbon, Oxygen, and Iron impurities).

The algorithms were tested under 5% Monte Carlo noise, with performance metrics including relative reconstruction error, structural correlation coefficients, and computational execution time. Sensitivity analyses were also conducted regarding the number of bolometer channels and the resolution of the reconstruction grid.

Key Contributions

• First Application of MLEM to IRVB: This work applies the MLEM algorithm to IRVB data reconstruction for the first time, evaluating its suitability against established regularization methods.
• Comprehensive Comparative Framework: The study provides the first side-by-side comparison of MFI, PTR-1, PTR-2, and MLEM specifically for the IRVB viewing geometry, which differs from conventional bolometer arrays by offering a denser, three-dimensional response matrix.
• Systematic Sensitivity Analysis: The paper quantifies how reconstruction accuracy and computational cost scale with varying bolometer channel counts (4×4 to 20×20) and reconstruction grid resolutions (11×11 to 41×41).
• Forward Modelling Validation: The work details the construction of synthetic phantoms and the forward modelling process, validating the inversion methods against expected emissivity profiles relevant to limiter and divertor operations.

Results

• Accuracy: The Minimum Fisher Information (MFI) method consistently achieved the lowest average relative reconstruction error (15.10%) across all phantoms. It demonstrated superior performance for compact and asymmetric structures, such as the double-null divertor profile (16.95% error vs. 43.05% for PTR-2). PTR-2 was the closest competitor with an average error of 18.72%. PTR-1 performed well for smooth, core-peaked profiles but degraded significantly for complex geometries (e.g., 53.65% error for double-null). MLEM exhibited the highest average error (31.09%) and failed to accurately reconstruct edge-dominated impurity profiles, showing a structural correlation coefficient of only 0.087 for the impurity phantom.
• Computational Performance: MLEM was the fastest method (0.36 s average) due to its simple element-wise updates, despite requiring up to 200 iterations. PTR-2 followed at 0.62 s, offering a closed-form solution. MFI was the most computationally demanding (3.02 s average) due to its iterative reweighted least squares (IRLS) inner loop.
• Sensitivity: Increasing the bolometer channel count generally reduced errors for regularization-based methods, though MFI's error increased for the impurity phantom at high channel counts due to noise amplification at the boundary. Regarding grid resolution, a 25×25 grid was identified as the saturation point for the 9×9 camera configuration; finer grids increased the problem's underdetermined nature without improving accuracy, while significantly increasing computation time, particularly for MFI.

Significance and Claims
The paper claims that the choice of reconstruction algorithm involves a practical trade-off between accuracy, numerical stability, and computational speed, dependent on the intended application:

• For Offline/Inter-shot Analysis: MFI is recommended as the optimal choice for high-fidelity reconstruction, particularly when analyzing complex, asymmetric, or compact emission structures where preserving peak features is critical.
• For Near-Real-Time Applications: PTR-2 is identified as the preferred algorithm. It offers the best balance, providing reconstruction fidelity within 3–5 percentage points of MFI while maintaining a computation time below 0.7 seconds, making it suitable for cycle times on the order of 1 second.
• Limitations: The study notes that PTR-1 is not recommended for diagnostically demanding configurations involving non-Gaussian emissions, and MLEM, while fast, is unsuitable for edge-dominated profiles due to poor fidelity and noise amplification.

The authors conclude that their framework provides a validated workflow for IRVB geometry construction and algorithm selection, emphasizing that while MFI offers the highest accuracy, PTR-2 represents the optimal solution for real-time or near-real-time usage in current tokamak operations. Future work is proposed to incorporate 3D modeling of viewing volumes and comparisons with machine learning approaches.