🔭 astrophysics

Extending Galactic foreground emission with neural networks

This paper introduces a Cycle-GAN-based approach that leverages Planck thermal dust and HI4PI HI data to accurately simulate Carbon Monoxide (CO) emission, successfully reproducing its statistical properties to improve current models in poorly observed high-Galactic latitude regions.

Original authors: Giuseppe Puglisi, Avinash Anand, Marina Migliaccio

Published 2026-04-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Giuseppe Puglisi, Avinash Anand, Marina Migliaccio

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

Imagine you are trying to listen to a very faint, beautiful song playing in a distant room (the Cosmic Microwave Background, or the echo of the Big Bang). But, the room is filled with loud, chaotic noise from a party next door (our own Galaxy). This "party noise" comes from dust, gas, and stars, and it drowns out the song you want to hear.

To hear the song clearly, astronomers need to map out exactly what the party noise looks like so they can subtract it. One specific type of noise is Carbon Monoxide (CO) gas. It's like a specific instrument in the party band that is hard to hear in certain parts of the room because the signal is too weak or the equipment is too noisy.

Here is how this paper solves that problem using a clever trick with Artificial Intelligence.

The Problem: The "Blind Spot"

Astronomers have great maps of the "loud" parts of the galaxy (near the center), but the "quiet" parts (far from the center, at high latitudes) are a mess. The data there is so full of static (noise) that it's hard to tell where the gas actually is.

Previously, scientists tried to guess what the gas looked like in these quiet areas by using simple math models. But these models were like stamping a cookie cutter: they created shapes that looked okay, but they didn't match the real, messy, complex patterns of the universe. They were uncorrelated with the other things we could see, like dust or atomic hydrogen.

The Solution: The "Artistic Translator"

The authors used a type of AI called a Cycle-GAN (Cycle-Generative Adversarial Network). Think of this not as a calculator, but as a master art student and a strict art critic working together.

The Art Student (The Generator): This AI looks at clear, high-quality pictures of Thermal Dust and Atomic Hydrogen (which we can see easily everywhere). It tries to "paint" a picture of what the invisible Carbon Monoxide gas should look like in those same spots.
The Art Critic (The Discriminator): This AI looks at the student's painting and compares it to real, high-quality photos of Carbon Monoxide (from the noisy but clear regions). It yells, "That doesn't look real! The texture is wrong!" or "Good job, that looks like the real thing!"
The Cycle (The Safety Net): To make sure the student isn't just making random art, they have a rule: If the student turns a Dust map into a CO map, and then turns that CO map back into a Dust map, it must look exactly like the original Dust map. This forces the AI to learn the true relationship between the materials, not just random patterns.

The Training Process

The team fed this AI thousands of "tiles" (small 3x3 degree patches) of the sky.

Inputs: Clear maps of Dust and Hydrogen.
Targets: The "noisy" but real maps of Carbon Monoxide (only in areas where the signal was strong enough to trust).

The AI learned that where there is a specific swirl of dust, there is usually a specific swirl of CO gas. It learned the "grammar" of the galaxy.

The Results: Filling in the Blanks

Once trained, the AI was sent to the "quiet" parts of the galaxy where we had no good data.

The Old Way: The old models looked like static-filled TV screens or simple blobs.
The New Way: The AI generated maps that looked like filaments and clouds, just like the real dust and hydrogen. It recreated the complex, stringy structures of the galaxy.

They tested this by checking the "fingerprint" of the images (mathematical statistics). The AI's fake maps had the exact same statistical fingerprints as the real maps. It wasn't just guessing; it had learned the physics of how the gas behaves.

Why This Matters

This is a game-changer for two reasons:

Better Cosmology: By creating a perfect map of the "party noise" (Galactic foregrounds), astronomers can subtract it more accurately, allowing them to hear the "song" of the Big Bang (CMB) much more clearly.
New Discovery: It allows us to "see" Carbon Monoxide in parts of the galaxy we couldn't observe before, helping us understand how gas clouds form in the quiet, dark corners of the Milky Way.

In short: The authors taught an AI to look at the "skeleton" of the galaxy (dust and hydrogen) and imagine the "muscle" (Carbon Monoxide) that goes with it, filling in the missing pieces of the puzzle with stunning accuracy.

1. Problem Statement

The Cosmic Microwave Background (CMB) is contaminated by Galactic foreground emissions, particularly Carbon Monoxide (CO) molecular line emissions. Accurate characterization of these foregrounds is essential for isolating cosmological signals, especially for future CMB polarization experiments like the Simons Observatory (SO).

Key Challenges:

Data Scarcity at High Latitudes: Existing CO surveys (e.g., Dame et al., 2001) focus on the Galactic mid-plane. Observations at high Galactic latitudes ( $|b| > 30^\circ$ ) are often noise-dominated or non-existent due to the intrinsically low emissivity and the long integration times required for detection.
Limitations of Current Models: Previous modeling approaches (e.g., Puglisi et al., 2017; MCMole3D) rely on physical assumptions or extrapolations from low-latitude data. These models often fail to capture the complex, non-linear morphological correlations between CO and other tracers (like thermal dust and atomic hydrogen) and may be heavily contaminated by instrumental noise in the templates they use.
Polarization Contamination: Neglecting polarized CO emission (even at low fractional levels $<1\%$ ) can contaminate future CMB polarization measurements.

2. Methodology

The authors propose a novel approach using Cycle Generative Adversarial Networks (Cycle-GANs) to synthesize realistic CO emission maps by learning the non-linear relationships between CO and other multi-wavelength tracers.

Data Sources

Input Domains (Source):
- Thermal Dust: Planck GNILC component-separated map at 857 GHz (highest Signal-to-Noise Ratio, SNR).
- Atomic Hydrogen (HI): HI4PI survey N(HI) column density map.
Target Domain (Ground Truth):
- CO Emission: Planck Type-2 templates for rotational transitions $J:1-0$ and $J:2-1$ .
Data Selection & Preprocessing:
- SNR Filtering: Training targets were restricted to regions with SNR > 8 to ensure robustness against noise.
- Tiling: Full-sky maps were segmented into $3 \times 3$ deg $^2$ tiles ( $128 \times 128$ pixels).
- Normalization: Maps were log-transformed and min-max normalized to the range $[-1, 1]$ .
- Augmentation: Data was augmented via flipping to increase diversity.
- Training Strategy: A hybrid approach was used. Initial training used paired data (limited by SNR > 8 regions) to learn astrophysical correlations. The network then switched to unpaired training to generalize features across optically thin and thick regimes, utilizing the full sky for inputs.

Network Architecture

Framework: Cycle-GAN (Zhu et al., 2020) allows for image-to-image translation without paired examples.
Components:
- Generators ( $G_{AB}, G_{BA}$ ): ResNet-based Encoder-Decoder structures with 3 convolutional blocks, 6 Residual blocks, and 3 transposed convolutional layers. Instance Normalization and Tanh activation are used.
- Discriminators ( $D_A, D_B$ ): PatchGAN architecture ( $14 \times 14$ grid) to enforce high-frequency details.
Loss Functions:
- Adversarial Loss: Mean-Squared Error (MSE) instead of Cross-Entropy for stability.
- Cycle-Consistency Loss ( $L_{cyc}$ ): Ensures $A \to B \to A \approx A$ .
- Identity Loss ( $L_{id}$ ): Ensures if an image is already in the target domain, it remains unchanged.
- Hyperparameters: $\lambda_{cyc} = 10$ (increased to 20 during unpaired training) and $\lambda_{id} = 5$ .

3. Key Contributions

First Application of Cycle-GAN to Galactic CO: Demonstrates the feasibility of using unsupervised deep learning to synthesize CO emission maps from dust and HI tracers.
Overcoming Data Limitations: Successfully generates CO maps in high-Galactic latitude regions where direct observations are noise-dominated or missing, without relying on physical extrapolation models that ignore morphological correlations.
Robust Training Strategy: The transition from paired to unpaired training allows the model to learn complex, non-linear correlations while generalizing to regions outside the high-SNR training set.
Open Science: The code and trained weights are made publicly available to ensure reproducibility.

4. Results

The performance of the generated maps was evaluated using statistical metrics on a held-out test set and high-latitude regions.

Visual Inspection: The generated maps successfully reproduce filamentary structures and point-source features observed in the ground truth, closely matching the morphology of the input dust and HI maps.
Angular Power Spectra:
- The power spectra of the synthesized CO maps follow the same scaling laws as the ground truth (real CO data).
- The median power spectrum of the mock data overlaps with the ground truth within the $2\sigma$ excursion range across angular scales from $0.5^\circ$ to $3'$ .
- The model correctly reproduces the CO spectral scaling, distinct from the input dust and HI spectra.
Minkowski Functionals (MFs):
- MFs ( $V_0, V_1, V_2$ ) were used to assess non-Gaussianity and higher-order statistics (area, perimeter, connectivity).
- The MFs of the generated maps are consistent with the ground truth within the $2\sigma$ interval, confirming that the model learns the complex topological properties of the molecular clouds.
High-Latitude Generalization:
- When applied to high-latitude regions ( $|b| > 30^\circ$ ) not present in the training set, the Cycle-GAN produced smooth, diffuse, filamentary structures.
- Comparison with Pysm3/MCMole3D: Unlike the Pysm3 model (based on noisy Planck Type-1 maps), which showed noise-dominated, compact features, the Cycle-GAN output preserved the diffuse morphology seen in dust/HI, effectively "denoising" the prediction by relying on high-SNR input tracers.

5. Significance and Future Outlook

CMB Science: This work provides a transformative tool for creating synthetic CO foreground templates. This is critical for component separation in future CMB experiments (e.g., Simons Observatory, CMB-S4), particularly for regions where current data is insufficient.
Methodological Shift: It validates the use of deep generative models to overcome observational limitations in astrophysics, moving beyond simple physical extrapolations to data-driven morphological synthesis.
Future Work: The authors plan to develop a full Galactic CO emission model for high latitudes based on these results, which will be integrated into standard sky simulation packages (like Pysm3) to improve foreground cleaning pipelines.

In conclusion, the paper establishes that Cycle-GANs can effectively learn the complex, non-linear mapping between thermal dust/HI and CO emissions, generating statistically accurate and morphologically realistic CO maps even in data-sparse high-Galactic latitude regions.