🔭 astrophysics

Denoising Interferometric Observations Using Visibilities-Informed Neural Networks

This paper introduces VIREO, a machine learning-based method that incorporates interferometric point spread functions into a U-Net architecture to effectively denoise astronomical observations, outperforming traditional cleaning techniques and demonstrating general applicability for future interferometric data analysis.

Original authors: Jason P. Terry, Cassandra Hall, Sergei Gleyzer

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jason P. Terry, Cassandra Hall, Sergei Gleyzer

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Cleaning Up a Blurry Photo

Imagine you are trying to take a photo of a beautiful, intricate sandcastle on a beach. However, the camera lens is dirty, the wind is blowing sand everywhere, and the lighting is poor. The resulting photo is grainy, blurry, and hard to make out.

In the world of astronomy, scientists use giant arrays of telescopes (like the future Square Kilometer Array or the current ALMA) to take "photos" of baby planets forming in disks of gas and dust around stars. But just like your beach photo, these telescope images are often full of "static" (noise) and blur. This makes it very hard to see the tiny details, like the rings where planets are forming or the spiral arms created by their gravity.

This paper introduces a new tool called VIREO (Visibilities-Informed Reconstruction for Enhanced Observations). Think of VIREO as a super-smart photo editor that doesn't just guess how to clean the picture; it knows exactly how the camera messed it up in the first place.

The Problem: The "Dirty" Image

When telescopes combine their data, they don't produce a perfect picture immediately. They produce a "dirty" image.

The Noise: It's like trying to listen to a whisper in a crowded stadium. The signal (the planet) is there, but the background noise (the crowd) drowns it out.
The Blur: The telescope has a specific "fingerprint" of blur called a Point Spread Function (PSF). It's like looking at a light through a foggy window; the light spreads out, making it hard to tell where the edges are.

Traditional methods try to clean this up, but they often act like a generic "sharpen" filter. They might remove the noise, but they also accidentally blur out the tiny, important details of the sandcastle.

The Solution: VIREO's "Magic Glasses"

The authors created a machine learning model (a type of AI) to fix these images. But here is the secret sauce: VIREO doesn't just look at the blurry photo; it also looks at the "fingerprint" of the blur.

The Two Inputs: Imagine you are trying to restore a torn painting.
- Standard AI: Only looks at the torn painting and guesses what the rest should look like.
- VIREO: Looks at the torn painting AND holds a map that shows exactly how the tear happened and how the paint spread.
- In the paper, this "map" is the PSF (the Point Spread Function). VIREO feeds this map into the AI's brain so it understands exactly how the telescope distorted the image.
The Training: The team didn't just throw random pictures at the AI. They created nearly 1,500 fake "sandcastles" (simulations of protoplanetary discs) using supercomputers. They then intentionally "dirtyed" these fake images with noise and blur that matched real telescope data. They taught the AI to turn the dirty versions back into the clean originals.
The Result: When VIREO processes a real image, it uses that "map" of the blur to surgically remove the noise without smearing the delicate details.

What Did They Find?

The paper compares VIREO against two other methods:

The "Old Way" (CASA): The standard software astronomers use today. It's like using a basic photo editor. It helps, but it leaves some noise and sometimes smears the edges.
The "Blind AI" (PSF-Ignorant): An AI that tries to clean the image but doesn't know about the telescope's specific blur fingerprint. It does better than the old way, but it still makes mistakes.
VIREO (The Winner): Because it knows the "fingerprint" of the blur, it produces the cleanest images.
- Less Noise: The background static disappears almost completely.
- Sharper Details: The rings and gaps where planets hide become much clearer.
- Better Science: When they used the cleaned images to count how many planets were in the simulation, VIREO was the most accurate. It made fewer mistakes than the other methods.

Real-World Test: The "HL Tau" Photo

To prove it works on real data, the team applied VIREO to famous, existing photos of baby planets (like HL Tau and AS 209) taken by ALMA.

Before: The photos were beautiful but had a hazy background.
After VIREO: The background became pitch black, making the rings and gaps pop out with incredible contrast. It was like turning off the lights in a room to see a glowing neon sign more clearly.

Why Does This Matter?

The Square Kilometer Array (SKAO) is about to start taking thousands of these pictures. Some of them will be very noisy because the telescopes will be looking at faint objects quickly.

Speed: VIREO cleans an image in a fraction of a second (0.07 seconds), whereas the traditional method takes nearly a minute.
Efficiency: Because VIREO can clean up "noisy" images so well, astronomers might not need to stare at the sky for days to get a perfect picture. They can take shorter, noisier snapshots and use VIREO to clean them up, saving huge amounts of telescope time.

In short: VIREO is a smart AI that uses a "cheat sheet" (the telescope's blur pattern) to remove static from astronomical photos, revealing the hidden baby planets that were previously invisible in the noise.

Technical Summary: Denoising Interferometric Observations Using Visibilities-Informed Neural Networks

Problem Statement
The upcoming Square Kilometer Array Observatory (SKAO) and next-generation Very Large Array (ngVLA) will generate vast datasets of long-wavelength astronomical observations, particularly of protoplanetary discs. However, interferometric observations over reasonable timescales (e.g., tens of hours) often result in noisy images that obscure critical substructures, such as spiral wakes and gaps, which are essential for understanding planet formation. Traditional cleaning software, such as CASA (Common Astronomy Software Applications), while effective, struggles with significant noise levels that can inhibit detailed analysis. Furthermore, existing machine learning approaches for denoising interferometric data have often ignored the specific observational characteristics, such as the Point Spread Function (PSF), that define the noise and resolution of the data.

Methodology
The authors propose VIREO (Visibilities-Informed Reconstruction for Enhanced Observations), a machine learning-based method designed to denoise interferometric data by explicitly incorporating the observation's PSF into the model architecture and training process.

Data Generation: The study utilizes 1,489 smoothed-particle hydrodynamics (SPH) simulations of gas-only protoplanetary discs containing 0 to 4 planets. These are processed through the MCFOST radiative transfer code to generate ground-truth visibility models at 12.5 GHz. Synthetic "dirty" observations are created by convolving these models with PSFs derived from a 133-dish SKAO-Mid array configuration, adding thermal noise, and simulating various integration times (10, 100, and 1000 hours) and resolutions (34 and 67 mas).
Model Architecture: The core of VIREO is a U-Net convolutional neural network. Unlike standard denoising models, VIREO accepts two inputs: the noisy dirty image and the corresponding PSF.
- Feature-wise Linear Modulation (FiLM): The key innovation is the integration of FiLM layers within the U-Net's up-sampling layers and bottleneck. These layers extract a "context vector" from the PSF, comprising geometric beam features (FWHM, eccentricity, position angle) and a convolutional encoding of the PSF itself.
- Modulation: This context vector generates scaling ( $\gamma$ ) and shifting ( $\beta$ ) parameters that modulate the feature maps, allowing the network to selectively adapt its processing based on the specific observational beam and noise characteristics.
- Loss Function: The training objective combines Mean Absolute Error (MAE), Multi-Scale Structural Similarity Index Measure (MS-SSIM), and an $l_1$ -starlet loss to preserve fine substructures. Crucially, the loss function includes a term that computes the MAE between the PSF-convolved clean target and the PSF-convolved reconstructed image, ensuring the output is consistent with the observational PSF.
Comparative Baselines: The authors compare VIREO against:
1. Traditional CASA multi-scale deconvolution.
2. An identical U-Net architecture trained without PSF inputs or PSF-specific loss (PSF-ignorant).
3. A downstream classification/regression model (Swin V2 Vision Transformer) trained to detect and count planets, used as a proxy for scientific utility.

Key Results

Image Quality: VIREO significantly outperforms both CASA and the PSF-ignorant model across all metrics. It removes background noise almost entirely while preserving and enhancing substructures like rings, gaps, and spiral wakes.
Quantitative Metrics:
- Noise Reduction: VIREO achieves the lowest off-source corner RMS noise across all integration times and resolutions.
- Flux Conservation: The model maintains flux accuracy better than other methods, with nearly 95% of pixels showing a flux error of less than 5%.
- Signal-to-Noise Ratio (SNR): VIREO yields the highest peak SNR across all datasets.
Downstream Analysis: When applied to a planet detection and counting task, images denoised by VIREO result in the highest Area Under the Curve (AUC) for classification and the lowest Mean Absolute Error (MAE) for regression. The confusion matrices indicate VIREO produces the fewest false positives and false negatives compared to other methods.
Real Data Application: When applied to archival ALMA data (HL Tau, AS 209, Elias 2-27), VIREO successfully reduces background noise and enhances the contrast of substructures (e.g., deepening gaps and sharpening ring peaks) without introducing significant artifacts, despite being trained on synthetic SKAO-Mid data.
Efficiency: A single VIREO denoising operation takes approximately 0.07 seconds, roughly 800 times faster than a standard CASA cleaning run (approx. 59 seconds).

Significance and Claims
The paper claims that VIREO represents a generalizable advancement for interferometric data analysis. By explicitly informing the neural network of the observation's PSF, the model learns to distinguish between true astrophysical substructure and artifacts introduced by the telescope's beam and noise properties.

The authors assert that this approach is particularly valuable for the SKAO, where data may be inherently noisier due to the scale of observations. They argue that VIREO enables robust scientific analysis of protoplanetary discs using shorter integration times (e.g., tens of hours) that would otherwise be too noisy for traditional methods, thereby increasing the efficiency and data output of future observatories. The successful application to ALMA data suggests that visibility-informed models are a viable, general technique for cleaning and analyzing interferometric observations across different facilities, provided they are trained on appropriate datasets.