Neural blind deconvolution to reconstruct high-resolution ground-based solar observations

This paper introduces a novel physics-informed neural network approach for ground-based solar image reconstruction that simultaneously estimates the atmospheric point spread function and the high-resolution intensity distribution, outperforming state-of-the-art methods in revealing small-scale solar features.

The Gist

Here is an explanation of the paper using simple language and creative analogies.

The Problem: The "Wobbly Window" Effect

Imagine you are trying to take a crystal-clear photo of a tiny, intricate snowflake through a window on a windy day. The glass is vibrating, and the air outside is turbulent. Even if you have the most expensive, high-powered camera in the world, your photo will come out blurry and wavy.

This is exactly the problem astronomers face when looking at the Sun from Earth.

• The Sun: A massive, dynamic star with tiny, fascinating details (like magnetic loops and tiny explosions) that we desperately want to see.
• The Atmosphere: Our Earth's atmosphere is like that wobbly, windy window. It distorts the light coming from the Sun, turning sharp details into a blurry mess. This is called "atmospheric seeing."

To fix this, astronomers use Adaptive Optics (like a super-fast, robotic mirror that tries to cancel out the wind) and they take thousands of photos in a split second (a "burst"). The idea is that while the wind changes every millisecond, the Sun doesn't. If you take enough photos, you can mathematically figure out what the Sun really looks like by averaging out the blur.

The Old Way: Guessing the Blur

For decades, scientists have used complex math to reverse-engineer the blur. They try to guess what the "blur filter" (called the Point Spread Function or PSF) looked like for every single photo, and then they try to subtract that blur to reveal the sharp image underneath.

Think of it like trying to un-mix a smoothie. You know the ingredients (the Sun), but you don't know exactly how the blender (the atmosphere) chopped them up. The old methods try to guess the blender's settings based on a few rules. Sometimes they get it right, but often they leave behind "artifacts" (weird digital noise) or miss tiny details because their guesses about the blender were too rigid.

The New Way: Neural Blind Deconvolution (NeuralBD)

This paper introduces a new, smarter way to un-blur these solar photos. The authors, led by C. Schirninger, created a method called NeuralBD.

Here is how it works, using a simple analogy:

1. The "Infinite Canvas" Artist

Instead of treating the image as a grid of fixed pixels (like a digital photo), the new method uses a Neural Network (a type of AI) that acts like a master painter with an "infinite canvas."

• Old way: You ask, "What color is pixel #450?"
• New way: You ask the AI, "If I look at any coordinate (x, y) on this canvas, what color should it be?"
  The AI learns to draw the Sun continuously, filling in the gaps between pixels with smooth, realistic details. This prevents the "blocky" artifacts common in old methods.

2. The "Blind" Detective

The method is called "Blind Deconvolution" because the AI doesn't know what the blur looks like beforehand. It has to figure out two things at the same time:

1. The Real Image: What does the Sun actually look like?
2. The Blur: What did the atmosphere do to mess it up?

Imagine you are a detective trying to solve a crime. You have a blurry security camera photo.

• Old method: You assume the camera lens was dirty in a specific, predictable way.
• NeuralBD method: You let the AI imagine every possible way the lens could be dirty. It tries to draw the criminal (the Sun) and the dirty lens simultaneously. It keeps adjusting its drawing until the "blurry version" of its drawing matches the actual blurry photo you have. When they match perfectly, the AI has successfully "un-blurred" the image.

3. No Training Data Needed

Most AI needs to be trained on thousands of examples (e.g., showing it 1,000 blurry photos and 1,000 sharp photos so it learns the pattern).

• NeuralBD is different: It doesn't need a library of examples. It solves the math problem from scratch for each specific observation. It's like a chef who doesn't need a recipe book; they just taste the soup and adjust the spices until it's perfect. This makes it incredibly flexible and able to work on any telescope.

The Results: Sharper Than Ever

The authors tested this method in three ways:

1. Simulation: They created a fake Sun on a computer, blurred it artificially, and asked NeuralBD to fix it. It did a better job than any previous method, recovering details that were thought to be lost forever.
2. GREGOR Telescope: They used real photos from a 1.5-meter telescope in Spain. NeuralBD revealed tiny, bright spots on the Sun that were invisible in the standard "un-blurred" photos.
3. DKIST Telescope: They tested it on the world's largest solar telescope (4 meters in Hawaii). Even with this massive telescope, the atmosphere still blurs the image. NeuralBD managed to see details smaller than the telescope's theoretical limit, outperforming the standard tools used by the observatory.

Why This Matters

This is a big deal for solar physics.

• Seeing the Invisible: The Sun's magnetic activity drives space weather, which can disrupt satellites and power grids on Earth. To predict this, we need to see the smallest details of solar eruptions.
• Future Proof: As we build bigger telescopes (like the future European Solar Telescope), the atmosphere will still be the bottleneck. NeuralBD provides a way to squeeze every last bit of clarity out of the light, regardless of the telescope size.

In summary: The paper presents a new AI-powered "digital eraser" that doesn't just guess how to fix blurry solar photos; it learns to paint the Sun and the atmosphere's distortion simultaneously, revealing a level of detail we've never seen before from the ground.

Technical Summary

Here is a detailed technical summary of the paper "Neural blind deconvolution to reconstruct high-resolution ground-based solar observations."

1. Problem Statement

Ground-based solar telescopes (e.g., GREGOR, DKIST) offer high spatial, spectral, and temporal resolution for studying the solar atmosphere. However, Earth's turbulent atmosphere introduces phase errors, causing image degradation and blurring (seeing). While Adaptive Optics (AO) systems mitigate some distortions, residual atmospheric effects remain.

Current state-of-the-art post-facto reconstruction methods, such as Speckle Reconstruction and Multi-Frame Blind Deconvolution (MFBD/MOMFBD), rely on:

• Fourier Domain Processing: Transforming data back and forth between image and Fourier domains, which can introduce artifacts.
• Isoplanatic Assumption: Assuming the Point Spread Function (PSF) is constant across the field of view (FOV), limiting reconstruction to small patches.
• Parametric PSF Constraints: Estimating PSFs using a fixed set of basis functions (e.g., Karhunen-Loéve polynomials) and wavefront coefficients. This can fail to capture complex, non-diffraction-driven distortions (like jitter) and may not fully recover the true PSF.
• Computational Cost & Artifacts: High computational demands and potential reconstruction artifacts due to sparse PSF information.

The paper aims to develop a novel reconstruction method that achieves unprecedented spatial resolution without relying on Fourier domain constraints or predefined PSF models, enabling the study of the smallest spatial scales from the photosphere to the chromosphere.

2. Methodology: NeuralBD (Neural Blind Deconvolution)

The authors propose NeuralBD, a solver-based approach using Physics-Informed Neural Networks (PINNs). Unlike traditional deep learning models that act as function approximators trained on large datasets, NeuralBD solves the image formation equation directly without a pre-existing training dataset.

Core Mechanism:
The method solves the image formation equation for a burst of $K$ short-exposure frames:
$$i_k(x, y) = o(x, y) * p_k(x, y) + n_k(x, y)$$
Where:

• $o(x, y)$ is the true object intensity distribution.
• $p_k(x, y)$ is the unknown PSF for frame $k$.
• $n_k(x, y)$ is noise.
• $*$ denotes convolution.

Key Technical Components:

1. Neural Representation of the Object:

   • A fully connected neural network (8 hidden layers, 512 neurons each) maps continuous coordinate points $(x, y)$ directly to intensity values $o(x, y)$.
   • Input Encoding: Uses Gaussian positional encoding (Fourier features) to map coordinates to sine/cosine functions, enabling the network to capture high-frequency details.
   • Output Activation: Uses an exponential function (base 10) to ensure positive intensity values.
   • Sampling: Uses stochastic sub-pixel sampling (perturbed grid cells) rather than fixed pixel centers to represent a continuous intensity distribution.

2. Learnable PSF Parameters:

   • Unlike MFBD, which constrains PSFs to wavefront coefficients, NeuralBD models the PSF as a grid of $N \times N$ free parameters (learnable values).
   • Constraints: PSFs are constrained only by positivity and normalization ($\sum PSF = 1$).
   • Initialization: Initialized from a Gaussian distribution, but the network is free to converge to any solution, allowing for complex, non-parametric PSF structures.

3. Optimization Process:

   • The network simultaneously optimizes the object representation $o(x, y)$ and the PSF parameters $p_k$.
   • Loss Function: Minimizes the Mean Squared Error (MSE) between the synthesized burst (convolution of estimated object and estimated PSFs) and the observed short-exposure burst.
   • Pre-training: The object model is pre-trained on the best single frame (lowest RMS) to accelerate convergence.

3. Key Contributions

• Image Domain Solver: Shifts the deconvolution problem from the Fourier domain to the image domain, avoiding Fourier transform artifacts.
• Unconstrained PSF Estimation: Removes the need for predefined basis functions or wavefront models, allowing the model to learn arbitrary PSF shapes, including spatial shifts and complex jitter.
• Data-Free Training: Does not require a large dataset of ground-truth images; the physical image equation serves as the training constraint.
• Simultaneous Estimation: Successfully estimates both the high-resolution object and the distorting PSFs in a single optimization loop.
• Scalability: Lays the foundation for future spatially varying PSFs (anisoplanatism) to reconstruct larger Fields of View (FOV).

4. Results

A. Synthetic Data (MURaM Simulation)

• Setup: Tested on 1024x1024 synthetic sunspot data degraded by 50 distinct PSFs and noise.
• Performance: NeuralBD significantly outperformed the baseline Richardson-Lucy (RL) deconvolution (which used known ground-truth PSFs) and the state-of-the-art torchmfbd method.
  • MSE: 0.0004 (NeuralBD) vs. 0.0067 (RL) and 0.0038 (torchmfbd).
  • SSIM: 0.96 (NeuralBD) vs. 0.49 (RL) and 0.57 (torchmfbd).
  • PSNR: 34.41 (NeuralBD) vs. 21.7 (RL) and 24.22 (torchmfbd).
• Visuals: NeuralBD recovered fine granulation and penumbral structures with higher fidelity and less noise than competitors. It also accurately reconstructed the ground-truth PSFs.

B. Real Observations (GREGOR & DKIST)

• GREGOR (1.5m): Applied to G-Band (430.7 nm) and Blue Continuum (450.6 nm) data.
  • NeuralBD revealed finer details in bright points and sub-arcsecond features that were blurred or unresolved in Speckle and torchmfbd reconstructions.
  • Power Spectral Density (PSD) analysis showed NeuralBD maintained high-frequency power comparable to or exceeding speckle methods in specific ranges.
• DKIST (4m): Applied to Visible Broadband Imager (VBI) data.
  • NeuralBD resolved fine penumbral structures better than Speckle and torchmfbd, with reduced noise amplification.
  • The method successfully estimated unique PSFs for different wavelengths, demonstrating its ability to handle chromatic effects naturally.

C. Uncertainty Estimation

• Running five independent models on the same data showed consistent large- and small-scale structures.
• Uncertainty maps (standard deviation) showed minimal variation, indicating the model converges to a stable, unique solution despite the high degree of freedom in PSF estimation.

5. Significance and Future Work

• Superior Resolution: NeuralBD demonstrates the ability to reconstruct small-scale solar features that exceed the performance of current state-of-the-art methods (Speckle, MFBD).
• Instrument Agnostic: The method requires no prior knowledge of telescope aperture or wavelength constraints, making it applicable to any ground-based broadband imaging instrument (e.g., GREGOR, DKIST, EST).
• Noise Mitigation: The PINN approach inherently provides smooth representations, effectively mitigating noise without explicit filtering.
• Future Directions:
  • Spatially Varying PSFs: Extending the method to handle anisoplanatism (varying PSFs across large FOVs) via mosaicing or pixel-wise PSF estimation.
  • Acceleration: Reducing the current reconstruction time (~7 hours on an A100 GPU) via patch-based sampling or meta-learning initialization.
  • Spectropolarimetry: Potential extension to polarimetric data for magnetic field studies.

In conclusion, NeuralBD represents a paradigm shift in solar image reconstruction, moving from constrained, Fourier-based parametric models to flexible, physics-informed neural solvers that achieve higher resolution and fidelity in ground-based solar observations.