HOLISMOKES XVII: Detecting strongly lensed SNe Ia from time series of multi-band LSST-like imaging data

This paper presents a deep learning pipeline based on a ConvLSTM2D architecture that effectively detects strongly lensed Type Ia supernovae in multi-band time-series imaging data, achieving high true-positive rates with low false-positive rates by the 7th to 9th observation to enable timely follow-up for upcoming surveys like LSST.

The Gist

Imagine the universe as a giant, cosmic funhouse mirror. Sometimes, a massive galaxy sits between us and a distant explosion (a supernova), bending the light like a lens. This creates a "strongly lensed supernova" (LSN), where we see the same explosion multiple times, appearing as separate dots of light scattered across the sky.

Finding these is like finding a needle in a haystack, but they are incredibly valuable. They act as cosmic stopwatches, helping scientists measure the expansion rate of the universe and solve mysteries about dark energy.

However, there's a catch: these events are rare, and the "mirror" (the lensing galaxy) often hides the explosion. By the time astronomers realize what they are looking at, the explosion might be fading, making it too late to study it properly. We need to spot them immediately.

Enter the authors of this paper: a team of astronomers who built a super-smart AI detective to find these lensed supernovae in real-time.

The Problem: Too Much Data, Too Fast

Imagine the Vera Rubin Observatory (a massive telescope in Chile) is like a security camera that takes a picture of the entire sky every few nights. It will generate millions of "alerts" every night—flashes of light that could be anything: a dying star, a distant galaxy, a glitch in the camera, or a lensed supernova.

Human astronomers cannot look at millions of pictures a night. They need a robot to do the screening. But the robot can't just look at one picture; it has to look at a movie of the sky over time to see how the light changes.

The Solution: The "Time-Traveling" AI

The team built a deep learning model called ConvLSTM2D. That's a mouthful, but here's the simple analogy:

• Standard AI (The Snapshot): Imagine a security guard who looks at a single photo of a street. They can tell you if a car is there, but they can't tell if the car is moving, speeding up, or slowing down.
• This AI (The Movie): This model is like a detective who watches a video clip of the street. It doesn't just see the car; it sees the pattern of the car's movement over time.

In astronomy terms:

1. Spatial (The "Where"): It looks at the shape of the image. Does it see one dot, or two dots that look like they belong to the same object?
2. Temporal (The "When"): It watches how the brightness changes over days and weeks. Does it follow the specific "heartbeat" of a Type Ia supernova?
3. Multi-Color (The "Rainbow"): It looks at the object through different colored filters (like red, blue, green). Lensing often makes distant objects look redder than nearby ones.

How They Trained the Detective

You can't train an AI on real data alone because we haven't found enough lensed supernovae yet. So, the team created a virtual reality training ground:

1. The Stage: They took real photos of the sky from the Subaru Telescope (HSC).
2. The Actors: They digitally "injected" fake supernovae into these photos.
   • The Good Guys (Positive Examples): They simulated lensed supernovae appearing in front of giant red galaxies.
   • The Bad Guys (Negative Examples): They filled the rest with "imposters": normal supernovae, variable stars, and random glitches that look like explosions but aren't lensed.
3. The Drill: They fed the AI thousands of these "movies" (time series) and asked it to guess: "Is this a lensed supernova or a fake?"

The Results: Getting Smarter with Time

The AI is designed to make a guess after every single new photo it receives.

• After 1 photo: It's a bit unsure (like guessing a mystery movie after the first scene).
• After 7 photos: It gets very good at spotting the pattern (60% accuracy with very few false alarms).
• After 9 photos: It becomes a master detective (over 70% accuracy).

Key Findings:

• Color Matters: The AI that looked at all colors (multi-band) was much better than the one looking at just one color. It's like solving a crime by looking at the suspect's clothes, shoes, and face, rather than just their shoes.
• The "Double" vs. "Quad" Problem: The AI is better at spotting systems where the supernova splits into four images (quads) rather than two (doubles). It's easier to see four distinct dots than two that might be blurry together.
• The Tricky Imposter: The hardest thing for the AI to distinguish was a normal supernova exploding inside a giant red galaxy. Sometimes, a lensed supernova looks exactly like this, confusing the AI.

Why This Matters for the Future

The Vera Rubin Observatory (LSST) will start taking pictures soon. It will take photos 5 to 10 times faster than the telescope used for this training.

The authors believe that with this faster "frame rate," their AI will become even sharper. It will be able to spot these cosmic miracles days or even weeks earlier than current methods.

The Bottom Line:
This paper is about building a smart, color-sensitive, time-watching AI that can sift through a mountain of astronomical data to find the rarest, most valuable explosions in the universe before they fade away. It's the difference between finding a needle in a haystack by looking at one photo of the hay, versus watching a movie of the haystack and spotting the needle as soon as it moves.

Technical Summary

Here is a detailed technical summary of the paper "HOLISMOKES XVII: Detecting strongly lensed SNe Ia from time series of multi-band LSST-like imaging data."

1. Problem Statement

Strongly lensed supernovae (LSNe) are rare but invaluable for cosmology, particularly for Time-Delay Cosmography (TDC) to measure the Hubble constant ($H_0$) and constrain dark energy. However, discovering them is challenging due to:

• Rarity: Only a handful have been confirmed to date.
• Identification Difficulty: Distinguishing LSNe from unlensed supernovae (SNe) and other transients (e.g., variable stars, AGNs) requires analyzing complex spatial and temporal signatures.
• Data Volume: Upcoming surveys like the Vera C. Rubin Observatory's LSST will generate massive alert streams ($O(10^7)$ per night), making manual identification impossible.
• Host Galaxy Ambiguity: In many cases, the host galaxy of an LSNe is either not lensed or too faint to be detected, removing a key visual cue for identification.
• Early Detection Need: To utilize LSNe for cosmography and progenitor studies, they must be identified before peak brightness to enable rapid follow-up.

The paper addresses the need for an automated, deep learning-based pipeline capable of detecting LSNe Ia early in their evolution using multi-band, multi-epoch imaging data.

2. Methodology

A. Data Construction and Simulation

The authors constructed a realistic dataset by combining real observations with synthetic injections, using Hyper Suprime-Cam (HSC) data as a proxy for LSST (similar depth and filter characteristics).

• Positive Class (LSNe Ia):

  • Lenses: Luminous Red Galaxies (LRGs) from the HSC Wide layer.
  • Injection: Mock Type Ia SNe were injected into LRG cutouts at various evolutionary phases.
  • Lensing Physics: A Singular Isothermal Ellipsoid (SIE) + shear model was used to generate lensing configurations (image positions, magnifications, time delays).
  • Constraints: Only systems with Einstein radii $\theta_E \geq 0.5''$ and sufficient brightness ($>5\sigma$) were selected. Both "doubles" (2 images) and "quads" (4 images) were included.
  • Hosts: Crucially, host galaxies were excluded from the mock LSNe to simulate the ~50% of cases where the host is undetectable or not lensed.

• Negative Classes (Non-LSNe):

  1. HSC Variables: Real, unclassified variable sources from the HSC Transient Survey (including potential SNe, variable stars, and bogus detections).
  2. Simulated Normal SNe Ia: Unlensed SNe injected into LRGs and spiral galaxies.
  3. Blanks: Real empty sky regions used to add realistic noise variations across epochs.

• Cadence Matching: To ensure the model learns temporal patterns rather than cadence artifacts, the synthetic time series were forced to match the specific, often irregular, observation schedule of the HSC Transient Survey (5–12 epochs over 6 months).

B. Model Architecture: ConvLSTM2D

The core of the pipeline is a Convolutional Long Short-Term Memory (ConvLSTM2D) network, designed to capture both spatial and temporal correlations simultaneously.

• Input: Time series of 2D image cutouts ($59 \times 59$pixels,$\sim 10'' \times 10''$) across 4 bands ($g, r, i, z$).
• Handling Asynchrony: Since observations in different bands are not synchronized, the model uses zero-padding for missing bands at each time step.
• Architecture:
  • Branch 1 (Spatial-Temporal): Processes image cutouts through three stacked ConvLSTM2D layers with max-pooling and batch normalization. This retains the 2D structure of memory states, allowing the model to learn spatial features (e.g., multiple image positions, arcs) evolving over time.
  • Branch 2 (Temporal): Processes the timestamps of observations using a standard LSTM to learn characteristic time scales.
  • Output: Concatenated features passed to a dense layer with a sigmoid activation, outputting a probability $P \in (0, 1)$ that the source is a lensed SN.
• Training Strategy: The model predicts after every observation in the time series, allowing performance to be evaluated as a function of the number of observations ($N_{obs}$).

3. Key Contributions

1. Multi-Band ConvLSTM2D for LSNe: This is the first application of a ConvLSTM2D architecture specifically tailored for multi-band time-series imaging of lensed supernovae, leveraging color information for early detection.
2. Realistic Simulation Pipeline: Unlike previous works relying solely on synthetic data, this study injects mock LSNe into real HSC images of galaxies and incorporates real variable sources from the HSC Transient Survey as negative examples.
3. Host-Independent Detection: The model is explicitly trained to detect LSNe without host galaxy light, addressing a major limitation in current discovery strategies where the host is often invisible.
4. Early Detection Capability: The framework demonstrates that high-confidence classification is possible after only a few observations, well before the supernova reaches peak brightness.

4. Key Results

• Performance Evolution:

  • Rapid Improvement: Classification accuracy improves significantly with each additional observation.
  • Metrics: At a strict False Positive Rate (FPR) of 0.01%:
    • True Positive Rate (TPR) reaches $\geq 60\%$ by the 7th observation (~17 days post-detection).
    • TPR exceeds $\geq 70\%$ by the 9th observation (~25 days).
  • Relaxed FPR: If the FPR is relaxed to 0.1%, TPR reaches $\sim 60\%$ as early as the 4th observation.
  • Saturation: Performance plateaus after $\sim 9$ observations, indicating sufficient information has been gathered.

• Multi-Band vs. Single-Band:

  • The multi-band model significantly outperforms single-band models, especially in the early stages (first few observations).
  • Multi-band data provides crucial color information (LSNe Ia are typically redder due to higher redshifts) and richer temporal context, which single-band models lack.

• Sources of Confusion (False Positives):

  • The primary source of false positives is unlensed SNe Ia in LRGs. These can mimic LSNe Ia when:
    • A doubly lensed image falls within the bright core of the lens galaxy, masking the lensing signature.
    • Fainter counter-images are missed due to poor cadence, making the system appear as a single point source.
  • Quads vs. Doubles: Quad systems are detected more effectively than doubles due to stronger spatial signatures.
  • Separation: Systems with larger image separations ($\theta_E > 1.0''$) are easier to detect than those with smaller separations.

5. Significance and Future Outlook

• LSST Readiness: While trained on HSC-like data, the methodology is directly transferable to LSST. The higher cadence of LSST (5–10 times that of HSC) is expected to further boost performance, enabling detection even earlier in the light curve.
• Real-Time Application: The pipeline is designed for real-time processing of alert streams, facilitating the rapid identification required for follow-up spectroscopy and time-delay measurements.
• Cosmological Impact: By enabling the discovery of a "gold sample" of tens to hundreds of LSNe, this method will help resolve the Hubble Tension and provide independent constraints on dark energy and spatial curvature.
• Future Work: The authors plan to incorporate microlensing effects, include lensed host galaxies in simulations, and explore more sophisticated architectures (e.g., Transformers) to handle asynchronous multi-band data more naturally.

In conclusion, this paper presents a robust, deep-learning-driven solution for the early detection of strongly lensed Type Ia supernovae, a critical step toward unlocking the full cosmological potential of upcoming time-domain surveys.