🔭 astrophysics

MALLORN: Many Artificial LSST Lightcurves based on Observations of Real Nuclear transients

This paper introduces MALLORN, a dataset of over 10,000 simulated LSST light curves derived from real observations of nuclear transients, designed to train and evaluate photometric classifiers for identifying tidal disruption events ahead of the Vera C. Rubin Observatory's full survey operations.

Original authors: Dylan Magill, Matt Nicholl, Vysakh Anilkumar, Sjoert van Velzen, Xinyue Sheng, Thai Son Mai, Hung Viet Tran, Ngoc Phu Doan, Thomas Moore, Shubham Srivastav, David R. Young, Charlotte R. Angus, Joshua

Published 2026-07-31

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Dylan Magill, Matt Nicholl, Vysakh Anilkumar, Sjoert van Velzen, Xinyue Sheng, Thai Son Mai, Hung Viet Tran, Ngoc Phu Doan, Thomas Moore, Shubham Srivastav, David R. Young, Charlotte R. Angus, Joshua Weston

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 the universe as a giant, cosmic fireworks show. Every night, stars explode, black holes eat nearby stars, and galaxies flicker with activity. For decades, astronomers have been trying to catch these fleeting flashes, but they've been like people with tiny flashlights in a stadium full of fireworks. Now, a new, massive telescope called the Vera C. Rubin Observatory is about to turn on a blinding spotlight that will reveal a hundred times more of these cosmic events than we've ever seen before.

However, there's a catch. While the telescope will spot millions of these "transients" (the fancy word for things that flash and fade), astronomers don't have enough giant telescopes with spectrographs (devices that break light into a rainbow to identify what something is made of) to check every single one. It's like having a million new suspects in a crime scene but only enough detectives to interview a few. To solve this, scientists need to become master detectives who can guess who the culprit is just by looking at the shape of the footprints (the light curve) left behind, without needing to interview the suspect directly. The biggest prize in this cosmic detective game is finding Tidal Disruption Events (TDEs)—moments when a supermassive black hole rips a star apart. These events are rare, hard to spot, and hold the secrets to how black holes grow, but they are often hidden in a crowd of look-alikes like exploding stars and active galactic nuclei.

Enter the MALLORN project, a clever new tool designed to train the next generation of cosmic detectives. The authors, led by Dylan Magill, realized that to teach computers to spot these rare black-hole feasts, they needed a massive practice field. They couldn't wait for the new telescope to start, nor could they rely on the few real examples they already had. So, they built a "simulator" that creates thousands of fake, but incredibly realistic, light curves based on real data from the Zwicky Transient Facility (ZTF).

Think of MALLORN as a high-tech "flight simulator" for astronomers. Just as a pilot practices in a simulator that mimics real weather and mechanical failures before flying a real plane, astronomers can now practice identifying TDEs using MALLORN's simulated data. The team took real observations of 64 TDEs, 727 nuclear supernovae, and 1,407 active galactic nuclei (AGN) and used a mathematical technique called Gaussian Process fitting to smooth out the data and fill in the gaps. They then "shrank" these objects to make them look like they were much farther away, simulating how they would appear to the new, more powerful LSST telescope.

The result is a dataset called MALLORN, which contains over 10,000 simulated light curves. These aren't just random numbers; they are carefully crafted to look exactly like what the LSST will see, complete with realistic noise, timing gaps, and color changes. The authors used a special "recipe" to translate the light from the ZTF's two main color filters into the six different colors the LSST will use, ensuring the fake data looks as authentic as possible.

To test if this approach works, the team launched the "MALLORN Astronomical Classification Challenge" on Kaggle, a platform where data scientists and the public compete to solve puzzles. The goal is simple: use the training data to build a computer algorithm that can correctly identify the TDEs hidden within the testing data. The paper explicitly notes that while they simulated the data, they did not claim to have discovered new TDEs; instead, they created a sandbox to improve the tools we will use when the real data arrives. They found that their method could successfully generate realistic light curves, but they also highlighted that some bands of light (like the red and infrared ones) might be too noisy to be very useful for finding faint TDEs, while the blue "u" band will likely be the most important for spotting them.

The paper concludes that while the simulation is a powerful step forward, it has limits. For instance, they couldn't simulate every possible type of rare transient because they didn't have real examples of them in their starting data. They also had to make some educated guesses about the temperatures of the objects to create variety, which might not perfectly match reality. However, by providing this massive, realistic dataset and a challenge to the community, MALLORN aims to sharpen our skills before the Vera C. Rubin Observatory even takes its first picture, ensuring that when the real cosmic fireworks show begins, we are ready to spot the most spectacular explosions of all.

Technical Summary: MALLORN – Many Artificial LSST Light Curves Based on Observations of Real Nuclear Transients

Problem Statement
The Vera C. Rubin Observatory's Legacy Survey of Space and Time (LSST) is projected to increase the discovery rate of transients by a factor of one hundred. However, spectroscopic resources are insufficient to follow up on every photometrically detected object. Consequently, the astronomical community must develop robust methods to prioritize targets for spectroscopy and analyze large samples based solely on photometry. This is particularly critical for identifying Tidal Disruption Events (TDEs), which are rare, physically significant events used to probe supermassive black hole (SMBH) masses and accretion physics. Current machine learning classifiers often struggle with contaminants like Active Galactic Nuclei (AGN) and Type Ia supernovae. Furthermore, previous classification challenges (e.g., PLAsTiCC, ELAsTiCC) relied heavily on simulated models that may not fully capture the complexity of real transients, risking poor performance when classifiers are deployed on real-world data.

Methodology
The authors present the MALLORN (Many Artificial LSST Light curves based on the Observations of Real Nuclear transients) dataset, a simulated photometric dataset designed to bridge the gap between real observations and LSST capabilities. The production pipeline involves the following steps:

Source Data Selection: The process begins with 11,190 nuclear transients from the Zwicky Transient Facility (ZTF) (2017–2023), filtered using the AMPEL Nuclear Filter. The final sample includes 64 confirmed TDEs, 727 nuclear supernovae (various types), and 1,407 AGN, all with spectroscopic labels.
Preprocessing: ZTF light curves undergo baseline correction (subtracting the pre-transient flux) and Galactic extinction correction using dust maps.
Continuous Light Curve Fitting:
- For supernovae and TDEs, a Gaussian Process (GP) fit is applied to the band with the most observations (typically $g$ or $r$ ). The kernel combines a constant, a Matérn 3/2 kernel (for smoothness), and a white noise kernel.
- For AGN, which exhibit stochastic variability, a Damped Random Walk (DRW) model is used within the GP framework to better capture the stochastic rises and falls.
- Fits with a Mean Squared Error (MSE) > 0.3 are discarded to ensure physical validity.
Luminosity and Redshift Rescaling: To simulate LSST's deeper limiting magnitude ( $m_{lim} \approx 24.7$ in $r$ -band vs. ZTF's 20.6), objects are virtually moved to higher redshifts. The distance is calculated such that the signal-to-noise ratio in the simulated LSST observation matches the observed ZTF signal-to-noise. Time dilation is applied to the light curve duration based on the new simulated redshift.
Multi-band Synthesis (SNCosmo): Since ZTF observes primarily in $g$ and $r$ , while LSST observes in six bands ( $u, g, r, i, z, y$ ), the authors use the SNCosmo package to generate empirically motivated Spectral Energy Distribution (SED) models. These models provide color evolution and $k$ -corrections to synthesize light curves for the missing bands ( $u, i, z, y$ ) based on the GP/DRW fit of the reference band.
Cadence and Noise Implementation: The continuous light curves are sampled using the LSST Baseline v5.0.0 cadence from the Rubin Survey Simulator. Random RA/Dec coordinates and peak dates are assigned within the survey footprint. Photometric errors are estimated based on the Rubin Survey Simulator's $5\sigma$ detection limits, and Gaussian scatter is added to mimic observational noise.
Quality Control: The final dataset underwent rigorous filtering, including checks for detection thresholds (minimum 10 points with $>5\sigma$ significance) and peak flux limits, resulting in 10,178 simulated light curves.

Key Contributions

The MALLORN Dataset: A dataset of 10,178 simulated LSST light curves derived from real ZTF observations, covering TDEs, various supernova types, and AGN.
Novel Simulation Method: A replicable framework that uses real transient observations as a base, applying GP/DRW fitting and SED modeling to simulate data for a new survey. This approach avoids the potential inaccuracies of purely analytic models.
The MALLORN Astronomical Classification Challenge: Hosted on Kaggle, this challenge provides a training set (30% of data with known spectroscopic types and true redshifts) and a testing set (70% with only photometric redshifts). The goal is to classify TDEs using the F1 score.
Open Tools: The authors released Google Colab notebooks detailing the data production process and the method for loading/de-extincting the data, lowering the barrier to entry for non-astronomers.

Results and Findings

Dataset Composition: The final dataset contains 10,178 light curves. The authors note that the simulation process naturally favors higher redshifts, as low-redshift objects are often already identifiable by existing methods.
Band Utility: Analysis suggests that $g$ , $r$ , and $i$ bands will be most useful for transient discovery due to frequent cadence and deep limits. The $u$ -band is identified as critical for TDE identification due to their blue nature, despite less frequent sampling. The $z$ and $y$ bands are found to be less useful for faint transient discovery due to shallower limits and infrequent sampling.
Detection Metrics: Applying detection metrics from Bricman et al. (2023) to the MALLORN TDE population, approximately 6.5% of TDEs pass the some_color metric, and 2.7% pass the some_color_pu (requiring a $u$ -band detection). Performance decreases significantly with increasing redshift.
Limitations: The authors acknowledge that the dataset excludes rare transients not present in the ZTF nuclear sample (e.g., kilonovae, SNe Ibn) and does not simulate failed photometric redshifts. Additionally, the random selection of temperatures for TDEs and AGN may not perfectly replicate the intrinsic luminosity-temperature correlations, though this was necessary to generate variety from limited source objects.

Significance
The paper posits that MALLORN is a vital tool for preparing the astronomical community for the LSST era. By providing a dataset grounded in real observations rather than purely theoretical models, it offers a more realistic testbed for photometric classifiers. The associated challenge aims to improve the collective ability to identify TDEs photometrically, thereby maximizing the scientific return of LSST given the constraints on spectroscopic follow-up. The authors emphasize that the methodology is highly replicable, allowing it to be adapted for future surveys (e.g., Roman High-Latitude Time-Domain Survey) by simply adjusting limiting magnitudes and cadence parameters.

More like this