🔭 astrophysics

Improved photometric redshift estimations through self-organising map-based data augmentation

This paper presents a Self-Organising Map-based data augmentation framework that significantly improves photometric redshift estimation accuracy, particularly for high-redshift galaxies, by identifying and filling spectroscopic gaps with simulated data to reduce systematic biases and catastrophic failures for future LSST surveys.

Original authors: Yun-Hao Zhang, Joe Zuntz, Irene Moskowitz, Eric Gawiser, Konrad Kuijken, Marika Asgari, Henk Hoekstra, Alex I. Malz, Ziang Yan, Tianqing Zhang, The LSST Dark Energy Science Collaboration

Published 2026-04-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yun-Hao Zhang, Joe Zuntz, Irene Moskowitz, Eric Gawiser, Konrad Kuijken, Marika Asgari, Henk Hoekstra, Alex I. Malz, Ziang Yan, Tianqing Zhang, The LSST Dark Energy Science Collaboration

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: Guessing the Distance of Stars

Imagine you are looking at a vast, dark forest at night. You see thousands of glowing fireflies (galaxies) scattered everywhere. Some are close and bright; others are far away and dim.

In astronomy, knowing how far away a galaxy is (its "redshift") is the most important piece of information. It tells us how fast the universe is expanding and how old the galaxy is.

The Gold Standard: The most accurate way to measure distance is to take a "spectrum" of the light (like taking a high-resolution photo of the firefly's specific color pattern). This is like using a laser rangefinder. It's accurate, but it takes a long time and is very expensive to do for millions of fireflies.
The Problem: We have telescopes (like the upcoming Rubin Observatory) that can take pictures of billions of galaxies quickly. But for these billions, we only have "blurry photos" (photometry), not the laser rangefinder data. We have to guess the distance based on how bright and what color they look.
The Risk: If we guess wrong, our entire map of the universe is wrong. The main reason we guess wrong is that our "training manual" (the few galaxies we did measure accurately) is missing huge sections of the forest. It has lots of close, bright fireflies, but very few of the faint, distant ones.

The Solution: A Smart "Fill-in-the-Blanks" Machine

The authors of this paper created a clever system to fix this missing data problem. They call it SOM-based Data Augmentation. Let's break down the analogy:

1. The Map (The Self-Organizing Map)

Imagine you have a giant, flat map of the forest. You want to organize every firefly on this map based on two things: how bright it is and what color it is.

Normally, you'd try to draw a 100-dimensional map, which is impossible for humans to visualize.
The authors used a computer algorithm called a Self-Organizing Map (SOM). Think of this as a magical 2D floor tile system. The computer automatically sorts every galaxy onto a specific tile based on its color and brightness.
The Discovery: When they looked at the tiles, they realized some tiles were packed with real, measured galaxies, but other tiles (especially the ones for faint, distant galaxies) were completely empty. These "empty tiles" are the danger zones where our guesses will fail.

2. The Fake Fireflies (Data Augmentation)

To fix the empty tiles, the team didn't just guess; they brought in a simulator.

They used a super-computer simulation (CosmoDC2) that generates millions of "fake" galaxies with realistic physics.
They took these fake galaxies and placed them onto the empty tiles on their map.
The Trick: They didn't just dump them randomly. They used a weighting system to make sure the fake galaxies looked exactly like the real ones in terms of distribution. It's like hiring actors to fill in the empty seats in a theater so the audience (the computer) thinks the theater is full.

3. The Training School

Now, the computer has a "Training Class" that is complete.

Before: The class had 100 students, but 50 of them were missing (the distant ones). The teacher (the AI) couldn't learn how to recognize the missing students.
After: The class has 100 real students + 50 carefully selected "fake" students filling the gaps. The teacher can now learn the rules for everyone.

The Results: Why It Matters

The team tested this method by simulating the first year (Y1) and the tenth year (Y10) of the Rubin Observatory's mission.

The "Before" Scenario: Without the fake fillers, the computer got very confused about the distant galaxies. It often made massive mistakes, saying a galaxy was close when it was actually far away (a "catastrophic failure").
The "After" Scenario: With the SOM-based augmentation:
- Mistakes dropped by half: The number of huge errors was cut by about 50%.
- Bias vanished: The system stopped consistently guessing "too close" or "too far."
- Confidence increased: The computer became much more reliable, even when looking at the faintest, most distant galaxies.

The Bottom Line

This paper is about teaching a computer to be a better astronomer by giving it a more complete textbook.

Instead of trying to measure every single galaxy perfectly (which is impossible), they used a smart map to find the holes in their knowledge and filled those holes with high-quality simulations. This ensures that when the Rubin Observatory starts taking pictures of the universe, we can trust the distance measurements enough to solve the biggest mysteries of the cosmos, like Dark Energy and Dark Matter.

In short: They built a better map, filled in the blank spots with smart simulations, and taught the AI to navigate the universe without getting lost.

1. Problem Statement

Accurate photometric redshift ( $z_{phot}$ ) estimation is critical for cosmological surveys like the Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST). However, current methods face significant challenges:

Spectroscopic Incompleteness: Training machine learning (ML) models for $z_{phot}$ requires spectroscopic redshifts ( $z_{spec}$ ), which are expensive to obtain. Existing spectroscopic surveys suffer from magnitude limits and uneven coverage in color-redshift space, leading to non-representative training sets.
Selection Biases: These gaps cause systematic biases, particularly at high redshifts ( $z \gtrsim 1.5$ ) and faint magnitudes, resulting in catastrophic failures and inaccurate conditional probability density functions (PDFs).
Model Mismatches: Traditional template-fitting methods rely on Stellar Population Synthesis (SPS) models that may not perfectly match real galaxy populations, while data-driven ML methods struggle to extrapolate beyond the specific distribution of their training data.

2. Methodology

The authors propose a framework using Self-Organizing Maps (SOMs) to identify gaps in spectroscopic coverage and perform targeted data augmentation using simulated galaxy catalogs.

A. Data Sources

Target Proxy (Application Dataset): The OpenUniverse2024 catalog, which uses differentiable physics-based models (Diffsky) to simulate realistic LSST-like observations (Y1 and Y10 depths). This serves as the "truth" for the photometric sample.
Augmentation Source: The CosmoDC2 catalog, derived from the Outer Rim N-body simulation using semi-analytic models (Galacticus/UniverseMachine). While realistic, it exhibits distinct SED (Spectral Energy Distribution) differences compared to OpenUniverse2024, simulating the systematic discrepancies between different simulations and real data.

B. The SOM-Based Pipeline

SOM Training: A 2D grid (SOM) is trained on the high-dimensional broadband SEDs of the OpenUniverse2024 Application dataset. This compresses the complex color-magnitude space into discrete cells (140×140 for Y1, 180×180 for Y10).
Degradation (Spectroscopic Simulation): The authors generate 501 realizations of synthetic spectroscopic datasets ("Degradation") by applying realistic selection cuts (magnitude, redshift, color) and probabilistic success rates to the Application dataset. This mimics the sparse, biased nature of real spectroscopic compilations.
Weighted Sampling & Augmentation:
- CosmoDC2 galaxies are mapped to the pre-trained SOM.
- Weighted Resampling: CosmoDC2 galaxies are re-weighted to match the cell occupancy of the Application dataset, correcting for global SED distribution differences.
- Targeted Augmentation: The algorithm identifies SOM cells that are empty or sparsely populated in the Degradation datasets. It then injects weighted CosmoDC2 galaxies into these specific cells to fill the gaps.
- Adaptive Scaling: The ratio of augmented to degraded galaxies is dynamically set based on the fraction of missing cells ( $f^*$ ), preventing the synthetic data from dominating the training set.
Combination Dataset: The final training set is a merger of the Degradation (realistic but sparse) and Augmentation (simulated but gap-filled) datasets.

C. Photometric Redshift Modeling

Algorithm: FlexZBoost (a hybrid of boosted decision trees and basis function decomposition) is used to estimate conditional PDFs $p(z|T)$ .
Training: Models are trained on the "Combination" datasets and validated against the Application datasets.
Metrics: Performance is evaluated using median scaled bias ( $\bar{\delta}_z$ ), Normalized Median Absolute Deviation (NMAD, $\sigma_z$ ), outlier fraction ( $f_o$ ), catastrophic failure rate ( $r_c$ ), and the divergence loss of Probability Integral Transform (PIT) histograms.

3. Key Contributions

SOM-Guided Targeting: Unlike previous augmentation methods that used simple color-magnitude boundaries, this approach uses SOMs to visualize and target specific high-dimensional gaps in the SED space, allowing for more precise filling of under-sampled regions.
Robustness to SED Mismatches: The study demonstrates that augmentation can significantly improve performance even when the augmentation source (CosmoDC2) has different underlying SED physics than the target (OpenUniverse2024).
Comprehensive Stress Testing: The generation of 501 independent realizations with varying noise, selection functions, and augmentation samples provides a rigorous statistical assessment of systematic uncertainties.
Dual-Depth Analysis: The framework is validated for both LSST Year 1 (Y1) and Year 10 (Y10) depths, addressing the increasing difficulty of high-redshift estimation as survey depth increases.

4. Results

High-Redshift Improvement: The most significant gains are observed at $z_{true} \gtrsim 1.5$ $z_{t r u e} ≳ 1.5$ .
- Catastrophic Failures: Reduced by a factor of $\sim 2$ in the Y1 scenario.
- Bias: Median scaled bias and outlier fractions decreased by factors of 2–3.
- PDF Fidelity: The divergence loss of conditional PDFs decreased by 25–50% at low redshifts and showed marked improvement in high-redshift tails, indicating better-calibrated uncertainty estimates.
Lens vs. Source Samples:
- Lens Samples (Bright, $z < 1.5$ ): Already well-represented by spectroscopic data; augmentation provided marginal improvements, confirming that bright galaxies are less affected by spectroscopic incompleteness.
- Source Samples (Faint, $z > 1.5$ ): Showed dramatic improvements. Without augmentation, models failed to extrapolate to high redshifts, resulting in massive catastrophic failure rates.
Reduced Scatter: The use of augmentation significantly reduced the variance in performance metrics across the 501 realizations, making the results less sensitive to specific spectroscopic selection functions.
Limitations: Some residual bias remains at very high redshifts ( $z > 2.0$ ) in the Y10 scenario due to persistent SED model differences between CosmoDC2 and OpenUniverse2024 that weighted resampling cannot fully reconcile.

5. Significance

This work provides a critical pathway for the LSST Dark Energy Science Collaboration (DESC) to achieve its cosmological goals:

Enabling Precision Cosmology: By mitigating catastrophic failures and systematic biases in $z_{phot}$ , the method ensures that weak gravitational lensing and large-scale structure analyses are not contaminated by redshift errors.
Scalability: The SOM-based approach is computationally efficient and adaptable to the petabyte-scale data expected from LSST.
Future Framework: The study establishes a foundation for future work on tomographic binning and ensemble redshift calibration, moving beyond point estimates to robust population-level statistics.
Validation: The results validate the use of synthetic catalogs for training ML models, provided that the training data is augmented to cover the full color-magnitude space of the target survey.

In summary, the paper demonstrates that SOM-based data augmentation is a powerful, necessary technique to overcome the inherent limitations of spectroscopic surveys, enabling the next generation of cosmological experiments to extract maximum scientific value from photometric data.