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

AI summary

Overview

Research area: Astrophysics (galaxy evolution) combined with machine learning — specifically unsupervised representation learning applied to Integral Field Spectroscopy (IFS) data from the MaNGA survey. arXiv category astro-ph.GA.

Technical level: Advanced. The conceptual goal is accessible, but the paper assumes familiarity with autoencoders, variational inference, LSTM recurrence, and spectroscopic diagnostics (emission lines, BPT diagrams).

Scope in one sentence: The paper builds and demonstrates an unsupervised Convolutional LSTM autoencoder framework that compresses spatially resolved galaxy spectra into a generalized latent representation and uses it to flag unusual galaxies, including active galactic nuclei (AGN).

What This Paper Is About

Integral Field Spectroscopy produces data cubes with two spatial dimensions and a spectral dimension, which yields enormous, high-dimensional datasets that are hard to explore by hand. The authors ask whether an unsupervised deep learning model can learn a compressed, general-purpose representation of both the spatial and spectroscopic structure of galaxies at once, without any labels. They then test whether that representation can surface scientifically interesting oddities — particularly unusual AGN — by measuring how badly the model reconstructs each galaxy.

Key Contributions

1. A spatio-spectroscopic representation learning framework. The authors introduce 2D Convolutional LSTM Autoencoder (2DConvLSTM-AE) and 2DConvLSTM-vAE architectures that operate on galaxy spectral cubes (X × Y × λ), treating a spectrum as a 1D sequence and a spatially resolved spectrum as a spatially correlated set of sequences. They note that 2DConvLSTMs had not previously been explored for IFS-based data.

2. A scalable data preparation pipeline for MaNGA. They convert raw MaNGA DRP data cubes into emission-line-only cubes sampling 19 optical emission lines (reported as 3800 Å < λ < 8000 Å in the abstract; the emission lines themselves span rest-frame 3272 Å for the OII doublet to 6733 Å for the SII doublet), with each line windowed at a bin width of ten corresponding to Δλ ~ 6 Å, giving 190 wavelength-wise dimensions, cropped centrally to 32 × 32 spaxels.

3. An unsupervised anomaly scoring scheme. For each galaxy they compute the mean absolute reconstruction error (MAE) between input and reconstructed cube as an "anomaly score," and visualize the latent space with UMAP after extracting the first 50 principal components of the latent vectors (corresponding to ~90% explained variance).

4. A demonstrated application to AGN. They assess the framework on 290 AGN selected from an overall sample of 406 AGN (Comerford et al. 2020, "C20") after a z < 0.08 cut, and use nearest-neighbour search in the representation space to retrieve galaxies with similar properties.

Main Findings

• The models learn a usable latent space. Latent embeddings from both the 2DConvLSTM-AE and 2DConvLSTM-vAE were extracted for ~9000 galaxies and visualized in a 3D UMAP space color-coded by anomaly score.

• Anomaly score distributions differ between the two models. Median anomaly scores are ~3000 for the 2DConvLSTM-AE and ~5000 for the 2DConvLSTM-vAE; the 90th percentile is ~12000 for the AE and ~20000 for the vAE.

• High-anomaly galaxies occupy distinct regions of the latent topology. Galaxies with low anomaly scores are spread across moderate-to-high values along the three UMAP dimensions, while high-scoring galaxies preferentially span the "wings" of the distribution with low UMAP1 and UMAP2 and high UMAP3 values.

• Most AGN sit near the low-anomaly locus, but a subset are anomalous. The 290 AGN generally fall within or close to the loci of low anomaly score samples, while some have high anomaly scores and occupy the same wings as other anomalous galaxies. Most of the anomalous AGN are radio-selected, with a handful that are X-ray and IR selected; most X-ray and broad-line selected AGN span regions populated predominantly by low to intermediate anomaly score samples.

• Highly anomalous AGN show diverse physical characteristics. Using the MaNGA Explorer tool, the authors report that these galaxies span disturbed morphology, substantial blue/purple emission in RGB images suggestive of co-existent rapid star formation, very strong emission in one or more emission lines, and AGN-like signatures in spaxel-wise BPT diagnostic plots ([OIII]/Hβ vs [NII]/Hα vs [SII]/Hα).

• One anomalous AGN is a known object of interest. Galaxy 8626-12704 is a "Blueberry" galaxy, previously studied by Paswan et al. (2022a, 2022b).

• Nearest-neighbour retrieval returns AGN-like galaxies. Using a kd-Tree method to compute pairwise Euclidean distances in UMAP space, the authors performed nearest-neighbour searches around anomalous AGN and found that the retrieved neighbours show emission line ratios in BPT plots indicative of an AGN.

• The method works without labels. The paper's headline claim is that anomalous galaxies, including AGN hosts, are identified in an unsupervised way and can be filtered and queried using the learned representation space.

Methodology in Plain English

The team started with the MaNGA survey's final data release (DR17). The summary catalog contains IFS observations of 11,273 objects, of which 10,010 are unique galaxies with reliable, good-quality data. They selected 9,043 galaxies with z < 0.08 so that the same optical emission lines fall within the MaNGA wavelength coverage. MaNGA galaxies were observed with different field-of-view layouts, from 19 fibers (diameter 12 arcsec) to 127 fibers (diameter 32 arcsec), and the raw cubes cover roughly 3600 Å ≲ λ ≲ 10300 Å with logarithmic wavelength spacing in flux units of 10⁻¹⁷ erg/s/cm²/Å/spaxel.

Instead of feeding raw spectra, they built emission-line-only cubes. For each galaxy, they used its redshift to locate the observed-frame wavelength of each of 19 emission lines, then took a ten-bin window around each center (Δλ ~ 6 Å) so the model would learn emission-line profile correlations rather than single central values. Because galaxies had different cube sizes (32 × 32 up to 127 × 127 spaxels), they cropped everything to 32 × 32 centered on the galaxy using RA and DEC. Input cubes are therefore 32 × 32 × 190. Fluxes were not normalized; models trained in native flux value space.

To help the model generalize across spatial and morphological variation, each cube was augmented three times with a random combination of horizontal flip, 90-degree rotation, Gaussian noise, and random spatial translation of up to 5 pixels, yielding ~36,000 augmented cubes. These were shuffled and split into ~30,000 training and ~6,000 test samples. Training used TensorFlow with a batch size of 16 cubes, a learning rate of 0.01, the Adagrad optimizer, and 30 epochs.

Architecturally, both models follow an encoder–bottleneck–decoder design. The encoder takes the 3D cube through two sets of wavelength-wise 2D convolutional blocks (each a series of λ Conv2D layers acting on 2D X × Y slices), then three 2DConvLSTM blocks with a downsampling factor of 2 in each. In the AE, the output is flattened into three fully connected layers whose central layer is the 512-dimensional bottleneck. In the vAE, two fully connected layers output the mean (μ_z) and log-variance (log σ_z²) of the latent distribution, and the latent vector is drawn randomly from a Gaussian. The decoder repeats the latent vector λ times, reshapes it, and passes it through three wavelength-distributed 2D transpose convolutional blocks that progressively upsample back to the input dimensionality. All but the final convolutional block use a linear activation followed by layer normalization; the final activation is ELU.

The AE loss is a batch-averaged sum of mean absolute errors computed both across spatial positions for each wavelength and across wavelengths for each spatial position. The vAE minimizes the sum of this reconstruction loss and the Kullback-Leibler divergence between the latent distribution and a unit normal. After training, reconstruction error served as the anomaly score, and the latent vectors were reduced to 50 principal components and projected with UMAP for visualization and nearest-neighbour querying.

Why This Matters

Impact on research. IFS surveys generate data whose dimensionality makes brute-force visual inspection impractical. This work shows that an unsupervised model can act as a compression and triage layer, turning cubes into compact vectors and ranking galaxies by how poorly the model understands them. It also extends a technique already common in imaging and 1D spectroscopy — autoencoder-based representation learning — into the combined spatial-spectroscopic domain. The paper reports that this combination had not previously been explored.

Real-world applications (as motivated by this work):

• Anomaly triage for large surveys: ranking millions of objects by reconstruction error to prioritize targets for scarce follow-up telescope time.
• Rare-object discovery: surfacing unusual galaxies such as the "Blueberry" galaxy 8626-12704 that are already subjects of dedicated study.
• Similarity search: using latent-space nearest-neighbour queries to find more objects like a given anomalous or scientifically interesting galaxy.
• Data compression and archival: storing compact latent representations instead of full high-dimensional data products.

Industry relevance. The core machinery — convolutional LSTM autoencoders, variational bottlenecks, reconstruction-error anomaly scoring, and latent-space similarity search — is the same toolkit used for anomaly detection in industrial sensor streams, video surveillance, medical time-series, and manufacturing quality control. The paper's contribution is a demonstration that these methods transfer to data where the "sequence" axis is spectral rather than temporal and where the anomalies are scientifically meaningful by construction.

Future Directions

1. Physical interpretation of the latent space. The paper visualizes latent structure but does not report which physical galaxy properties map onto individual latent dimensions; connecting axes to measurable quantities such as star formation rate, metallicity, or kinematics is a natural next step.

2. Systematic comparison of AE versus vAE. The two models produce different anomaly score distributions (median ~3000 versus ~5000; 90th percentile ~12000 versus ~20000), but the paper does not establish which is more effective for anomaly detection, nor does it report a quantitative benchmark of anomaly-detection performance.

3. Scaling and generalization. The work uses 9,043 galaxies with z < 0.08 and a fixed 32 × 32 cropping. Extending to larger samples, higher redshifts, other IFS surveys, or the full native spatial resolution remains open.

4. Follow-up confirmation of anomalous candidates. The paper demonstrates that nearest neighbours of anomalous AGN show AGN-like BPT ratios, but dedicated spectroscopic or multiwavelength follow-up of the newly flagged anomalous galaxies is needed to confirm what they are.

Target Audience

Astronomers working on galaxy evolution and IFS surveys who need scalable ways to explore large data sets; machine learning researchers interested in scientific applications of unsupervised sequence-and-image models; and survey scientists building anomaly detection or similarity-search pipelines. Readers without a spectroscopy background can follow the modeling narrative, but the interpretation of results assumes familiarity with emission lines, BPT diagnostics, and AGN selection methods.

Authors’ abstract

Integral Field Spectroscopy (IFS) surveys offer a unique new landscape in which to learn in both spatial and spectroscopic dimensions and could help uncover previously unknown insights into galaxy evolution. In this work, we demonstrate a new unsupervised deep learning framework using Convolutional Long-Short Term Memory Network Autoencoders to encode generalized feature representations across both spatial and spectroscopic dimensions spanning $19$ optical emission lines (3800A $&lt; λ&lt;$ 8000A) among a sample of $\sim 9000$ galaxies from the MaNGA IFS survey. As a demonstrative exercise, we assess our model on a sample of $290$ Active Galactic Nuclei (AGN) and highlight scientifically interesting characteristics of some highly anomalous AGN.