Spatio-Spectroscopic Representation Learning using Unsupervised Convolutional Long-Short Term Memory Networks

This paper presents an unsupervised deep learning framework using Convolutional Long-Short Term Memory Network Autoencoders to learn spatio-spectroscopic feature representations from MaNGA survey data, successfully identifying scientifically significant characteristics in anomalous Active Galactic Nuclei.

The Gist

Imagine you are a detective trying to understand a massive city called "The Galaxy." In the past, astronomers looked at this city through a single, blurry window, seeing only the total amount of light coming from the whole neighborhood. But thanks to new technology called Integral Field Spectroscopy (IFS), they can now look at the city street by street, block by block, and even analyze the "chemical fingerprint" of the air in every single house.

This creates a mountain of data. It's like trying to read every book in a library the size of a small country, all at once. Humans can't do that alone. That's where this paper comes in.

Here is the story of how the authors built a super-smart robot librarian to find the weird, interesting, and mysterious galaxies hiding in that mountain of data.

1. The Problem: Too Much Data, Too Many Dimensions

The researchers used data from the MaNGA survey, which took "3D movies" of about 9,000 galaxies.

• The "3D" part: Imagine a galaxy not as a flat picture, but as a cube.
  • X and Y: The left/right and up/down view (the shape of the galaxy).
  • Z (Depth): The spectrum. Instead of just seeing colors, they see the specific "notes" (wavelengths) of light coming from 19 different chemical elements (like Hydrogen or Oxygen).

Trying to find a weird galaxy in this 3D cube is like trying to find a specific, slightly different note in a symphony played by 9,000 different orchestras simultaneously.

2. The Solution: The "Time-Traveling" Robot Librarian

The authors built a special kind of Artificial Intelligence (AI) called a 2D Convolutional Long-Short Term Memory Network Autoencoder. That's a mouthful, so let's break it down with an analogy:

• The "Autoencoder" (The Compression Machine): Imagine you have a giant, messy pile of laundry (the raw galaxy data). You want to fit it all into a tiny suitcase (the "Latent Space") to carry it around. The AI acts as a master packer. It squishes the galaxy data down into a tiny, efficient code that still holds all the important information.
• The "Convolutional" part: This is like a scanner that looks at the shape of the galaxy (is it a spiral? a blob?) while packing.
• The "LSTM" (Long-Short Term Memory): This is the magic ingredient. Usually, AI looks at a picture and stops. But galaxies have "spectra," which are like a sequence of notes. An LSTM is like a musician who remembers the melody. It understands that the light from one part of the galaxy is connected to the light from the next part. It learns the relationships between the different chemical "notes" across the whole galaxy.

The Goal: The AI tries to take a galaxy, compress it into a tiny suitcase, and then unpack it to recreate the galaxy perfectly. If the AI is good, the unpacked galaxy looks exactly like the original.

3. The Detective Work: Finding the "Weirdos"

Here is the clever trick: The AI is trained to be boring.

The AI learns what a "normal" galaxy looks like by practicing on 9,000 of them. It gets really good at packing and unpacking normal galaxies.

• The Test: When they feed it a galaxy that is strange or anomalous (like one with a crazy black hole or a weird explosion), the AI gets confused. It tries to pack it, but it can't fit the "weirdness" into its standard suitcase.
• The Result: When it unpacks the galaxy, it looks messy or wrong. The difference between the original and the messy unpacked version is called the "Anomaly Score."
  • Low Score: "This galaxy is normal. I know exactly what to do with it."
  • High Score: "Whoa! This galaxy is weird! I don't know how to pack this!"

4. The Discovery: The "Blueberries" and the Black Holes

The researchers tested this system on a group of galaxies known to have Active Galactic Nuclei (AGN)—basically, galaxies with super-massive black holes eating stars and glowing brightly.

• The Findings: Most of these black-hole galaxies looked "normal" to the AI. But a few stood out with high anomaly scores.
• The "Blueberry" Galaxy: One of the weirdest galaxies they found was a "Blueberry" galaxy. It's a small, intense burst of star formation that looks like a bright blue berry. The AI flagged it as "super weird," and when the humans looked at it, they confirmed it was indeed a scientifically fascinating object that had recently been studied.
• The Search: Because the AI had compressed all the galaxies into a "map of weirdness," the researchers could ask: "Show me all the galaxies that look like this weird Blueberry." The AI instantly found other galaxies with similar strange traits, even if they were hidden deep in the data.

Summary

Think of this paper as building a smart filter for the universe.

1. They taught a robot to memorize what "normal" galaxies look like in 3D.
2. They let the robot scan thousands of galaxies.
3. Whenever the robot got confused and said, "I can't understand this one," they knew they had found something special.
4. This allowed them to find rare, weird galaxies (like those with crazy black holes or intense star birth) much faster than looking at them one by one.

It's a new way to explore the universe: instead of looking for what you expect, you let the computer tell you what is unexpected.

Technical Summary

1. Problem Statement

Integral Field Spectroscopy (IFS) surveys, such as MaNGA (Mapping Nearby Galaxies at Apache Point Observatory), provide rich 3D data cubes containing spatial ($X, Y$) and spectral ($\lambda$) information for thousands of galaxies. While this data allows astronomers to study complex physical processes like star formation and Active Galactic Nuclei (AGN) evolution, the sheer volume and high dimensionality of IFS data pose significant challenges for traditional analysis.

Existing deep learning approaches in astronomy have primarily focused on:

• Imaging data: Using Convolutional Autoencoders (AEs) for galaxy morphology.
• 1D Spectra: Using standard AEs for spectral characterization.
• Sequential data: Using LSTMs for time-series or 1D sequences.

However, there is a gap in unsupervised frameworks capable of simultaneously learning generalized feature representations across both spatial and spectroscopic dimensions (spatio-spectral relationships) to detect anomalies or unusual physical signatures in galaxy data without prior labeling.

2. Methodology

The authors propose a novel unsupervised deep learning framework utilizing 2D Convolutional Long-Short Term Memory (2DConvLSTM) networks integrated into Autoencoder architectures.

A. Data Preparation

• Dataset: The model was trained on 9,043 galaxies from the MaNGA survey (SDSS DR17) with redshift $z < 0.08$.
• Input Construction:
  • Raw spectral cubes were processed to extract 19 specific optical emission lines (ranging from [OII] to [SII], covering $\sim$3800Å to 8000Å).
  • For each galaxy, a wavelength window of $\sim$6Å was defined around each line center to capture line profile correlations rather than single values.
  • The resulting data cubes were cropped to a uniform $32 \times 32$ spatial grid centered on the galaxy.
  • Input Dimension: $32 \times 32 \times 190$ (Spatial $\times$ Spatial $\times$ 19 lines $\times$ 10 wavelength bins per line).
• Data Augmentation: To handle spatial and morphological variances, each cube was augmented 3 times using random horizontal flips, 90° rotations, Gaussian noise, and spatial translations (up to 5 pixels), resulting in $\sim$36,000 training samples.

B. Model Architecture

The authors developed two architectures using TensorFlow, both following an Encoder-Bottleneck-Decoder structure:

1. 2DConvLSTM-AE (Autoencoder): A deterministic model.
2. 2DConvLSTM-vAE (Variational Autoencoder): A probabilistic model.

Encoder:

• Input: 3D Spectral Cube ($X \times Y \times \lambda$).
• Conv2D$\lambda$ Blocks: Two sets of wavelength-wise 2D convolutional blocks acting on 2D spatial slices.
• 2DConvLSTM Blocks: Three 2D Convolutional LSTM blocks with a downsampling factor of 2 per block to capture spatio-temporal (spatio-spectral) dependencies.
• Bottleneck:
  • AE: Flattened features pass through FC layers to a latent vector $Z$ (512 dimensions).
  • vAE: Flattened features pass through two FC layers to predict mean ($\mu_z$) and log-variance ($\log \sigma^2_z$). $Z$ is sampled from a Gaussian distribution $N(\mu_z, \sigma_z)$.

Decoder:

• The latent vector $Z$ is repeated $\lambda$ times and reshaped.
• Passed through three 2D Transpose Convolutional blocks (Conv2D$^T_\lambda$) to progressively upsample features back to the original $32 \times 32 \times 190$ dimension.
• Activation: Linear activation with Layer Normalization in intermediate layers; ELU (Exponential Linear Unit) in the final layer.

C. Training Strategy

• Loss Functions:
  • AE: Mean Absolute Error (MAE) between input and reconstructed cubes ($L_{rec}$).
  • vAE: Sum of $L_{rec}$ and Kullback-Leibler (KL) Divergence loss to regularize the latent space.
• Hyperparameters: Batch size of 16, Learning rate of 0.01, Adagrad optimizer, trained for 30 epochs.
• Anomaly Scoring: Anomaly scores are calculated as the Mean Absolute Error (MAE) between the input and the reconstructed cube. High reconstruction error implies the model failed to learn the pattern, indicating an anomaly.

3. Key Contributions

1. Novel Architecture: First application of 2DConvLSTMs to unsupervised learning of spatio-spectroscopic galaxy data, effectively treating spectral lines as a sequence of spatial maps.
2. Unsupervised Anomaly Detection: Demonstrated the ability to identify scientifically interesting, anomalous galaxies (including AGN) without labeled training data, relying solely on reconstruction error.
3. Latent Space Exploration: Successfully mapped the high-dimensional spatio-spectral data into a 3D UMAP latent space, revealing distinct clusters and "wings" corresponding to anomalous objects.
4. Scientific Validation: Validated the model by identifying known AGN and discovering new candidates, including a "Blueberry" galaxy (8626-12704), confirming the physical relevance of the learned representations.

4. Results

• Latent Space Distribution:
  • The majority of galaxies with low anomaly scores clustered in the center of the UMAP space.
  • Galaxies with high anomaly scores (the "wings" of the distribution) were characterized by low UMAP1/UMAP2 values and preferentially high UMAP3 values.
• AGN Performance:
  • The model was tested on a subset of 290 AGN (from the C20 catalog).
  • While many AGN fell within the normal distribution, a significant subset of radio-selected AGN exhibited high anomaly scores, aligning with the "wings" of the distribution.
  • X-ray and broad-line selected AGN were more dispersed, often overlapping with low-to-intermediate anomaly samples.
• Case Studies:
  • Highly anomalous galaxies identified by the model showed disturbed morphologies, strong blue/purple emissions (indicating rapid star formation), and strong emission lines.
  • Nearest Neighbor (NN) Search: Using the latent space, the authors queried anomalous AGN and found neighbors that shared similar BPT diagnostic plots (indicating AGN activity), proving the model learned physically meaningful features.

5. Significance

This work establishes a powerful, data-driven pipeline for discovering rare or unusual astrophysical phenomena in large-scale spectroscopic surveys. By moving beyond traditional 1D spectral analysis or 2D imaging analysis, the 2DConvLSTM framework captures the complex interplay between spatial structure and spectral features.

The significance lies in:

• Scalability: The unsupervised nature allows the model to be applied to future massive IFS surveys (e.g., SDSS-V, DESI) without requiring pre-labeled anomaly datasets.
• Discovery Potential: It successfully filters for "needle in the haystack" objects, such as specific types of AGN or interacting galaxies, which might be missed by standard selection criteria.
• Methodological Advancement: It bridges the gap between sequential data modeling (LSTMs) and spatial data modeling (CNNs) specifically for 3D astronomical data cubes.