ZTF-SEDm Type Ia supernova sample for Twins Embedding spectrophotometric standardisation

arXiv:2512.07696 · astro-ph.CO · Submitted 2025-12-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Astrophysics Radio. Generated commentary on the latest astrophysics papers.

Vera: Today's paper: "ZTF-SEDm Type Ia supernova sample for Twins Embedding spectrophotometric standardisation".

Jocelyn: This paper presents the first application of the Twins Embedding (TE) spectrophotometric standardisation method to a large, heterogeneous dataset, addressing the challenge of using low-resolution spectra for precise cosmological measurements.

Vera: First, who's behind it and why it matters.

Title and authors: Vera: This paper, "ZTF-SEDm Type Ia supernova sample for Twins Embedding spectrophotometric standardisation," is essentially about improving how we calibrate SN Ia data when our spectral resolution isn't perfect.

Jocelyn: I think the title itself tells us exactly what the focus is: they are using a specific dataset from ZTF and applying Twins Embedding to achieve a level of standardisation that's typically hard to reach with lower quality spectra.

Subrahmanyan: From a theoretical standpoint, this work addresses the empirical procedure of deriving distances from SN Ia data, which inherently carries an accuracy uncertainty that affects our inferred cosmology.

Vera: Right, and the authors are doing this by building up a sample and then applying two main steps: first using SALT2 point 4 to interpolate light curve data to get flux calibration, and then using RBTL followed by Twins Embedding for the final correction.

Jocelyn: So they're not just looking at one method; they're combining spectral fitting with manifold learning to see if they can isolate the common variability across all SNe Ia more effectively than before.

Subrahmanyan: The implication here is that if you can get a robust standardisation method that minimizes astrophysical biases, we gain more reliable distance indicators for cosmological probes.

Vera: It seems the authors are focusing heavily on overcoming the unknown absolute flux calibration of SEDm spectra by comparing them to high-quality ZTF photometric data.

Jocelyn: That initial flux calibration step using SALT2 point 4 interpolation sounds like a very necessary bridge between the raw spectroscopy and usable photometric information for these one thousand eight hundred ninety-seven flux-calibrated spectra from one thousand six hundred seven SNe Ia.

The paper's summary: Vera: To summarize what they did, the paper first sets out to build this large homogeneous sample using the SEDm IFS data and then demonstrate that Twins Embedding can successfully provide high-precision standardisation even with limited spectral quality.

Jocelyn: They tackle a few specific hurdles, starting with flux calibration accuracy, which they achieve by interpolating light curve data and correcting for second-order polynomials to get percent level accuracy in the four thousand five hundred–seven thousand Å range.

Subrahmanyan: The methodology then moves into two distinct phases: first applying the Read Between The Lines method to extract a single representative spectrum at maximum light, and second using the Twins Embedding technique on those spectra.

Vera: That RBTL step is designed to capture the color term by ignoring strong absorption lines, which they argue makes it less prone to astrophysical bias compared to purely photometric methods.

Jocelyn: After that RBTL fitting, they estimate parameters like the color term and magnitude offset by minimizing a chi squared function while accounting for chromatic intrinsic dispersion and calibration uncertainty.

Subrahmanyan: The subsequent Twins Embedding stage uses an Isomap algorithm on these corrected spectra to capture variances not covered by the initial color and magnitude terms, which are then fed into a Gaussian Process to predict the remaining gray offsets.

Vera: The overall goal is to arrive at a complete correction formula, mu TE,i = -m i - beta RBTL times AV i - delta m GP(xi i), which aims to reduce the scatter further.

Jocelyn: However, the paper also notes a significant limitation: that for their specific ZTF sample, the residuals dispersion didn't improve much because of high instrumental error in that data.

The paper's improvements: Vera: The authors are proposing several refinements to this process to make it more robust, specifically focusing on how they handle the environmental and spectral features.

Jocelyn: One key improvement they suggest is an Adaptive Spectroscopic Decontamination Module where a deep learning model could predict and correct for residual host galaxy contamination based on local metrics like dDLR or stellar mass M host.

Subrahmanyan: That would be an interesting step because it moves the correction from a fixed subtraction to something dynamic, learning that specific spectral slopes are correlated with local environmental factors, which is vital for reducing astrophysical bias.

Vera: I also saw a suggestion to integrate Dynamic Weighting into the DTEM process, where the intrinsic spectral scatter function, eta(lambda), acts as a weighting mask.

Jocelyn: That would mean when correcting for phase evolution, if there are strong absorption lines that are poorly constrained by low-resolution SEDm, the AI automatically deweights those unstable wavelength bins during the calculation of the model spectrum.

Subrahmanyan: If you dynamically weight based on feature stability, you're ensuring that the resulting spectrum at maximum light is defined only by its most robust spectral domains, which should lead to a more reliable parameter estimation for things like AV.

Vera: And then there's the idea of Transfer Learning for the Isomap manifold learning process itself, adapting it to the specific feature set observed in ZTF.

Jocelyn: So instead of relying on pre-trained vectors from other datasets, if ZTF lacks certain spectral features used in training, this system would adaptively adjust those manifold parameters to account for that missing information and lower the residual dispersion.

Conclusion: Vera: So wrapping up the discussion on "ZTF-SEDm Type Ia supernova sample for Twins Embedding spectrophotometric standardisation," the paper confirms that spectroscopic methods can indeed yield low Hubble residual scatter, with their initial RBTL standardisation showing a value of zero point one five three mag in nMAD.

Jocelyn: And while the Twins Embedding method showed promise, they also noted that the residuals dispersion didn't improve much for their ZTF sample because of high instrumental error, especially when compared to data from SNfactory which achieved zero point zero nine seven mag.

Subrahmanyan: The implication is that when we can reduce the amplitude of astrophysical bias using RBTL standardisation, it confirms that spectroscopic approaches are superior to SALT photometric methods for achieving a cleaner fit, as the amplitude of the astrophysical bias was reduced to steps consistent with zero.

Vera: It’s a big win because it gives us an alternative path for calibrating SN Ia distances that isn't entirely dependent on photometric relations that have their own astrophysical uncertainties.

Jocelyn: We are excited to see how these proposed improvements, like the dynamic weighting and decontamination module, can handle the noise in real observational data to push those residual errors even lower.

Subrahmanyan: Ultimately, this paper demonstrates a pathway for using large spectroscopic samples to constrain cosmology with high-precision tools that are less susceptible to those photometric biases.

Vera: That’s what we've covered today regarding the ZTF-SEDm Type Ia supernova sample for Twins Embedding spectrophotometric standardisation. We’ll be back next time.

University of Claude Bernard Lyon 1 (IP2I Lyon/IN2P3, CNRS) · Lancaster University (Department of Physics) · The Oskar Klein Centre, Department of Physics · Lawrence Berkeley National Laboratories · Caltech Optical Observatories · University of Maryland · University of Washington · Deutsches Elektronen-Synchrotron (DES) · TANGO Consortium of Taiwan

astro-ph.CO

Submitted: 2025-12-08

Updated: 2025-12-08

Journal ref: A&A, 710, A315 (2026)

DOI: 10.1051/0004-6361/202558467

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: This paper presents the first application of the Twins Embedding (TE) spectrophotometric standardisation method to a large, heterogeneous dataset, addressing the challenge of using low-resolution

Key concepts

Twins Embedding (TE)
A manifold learning technique used on corrected spectra to capture variances not covered by initial color and magnitude terms. It is fed into a Gaussian Process to predict remaining gray offsets, aiming for a complete correction formula.
Read Between The Lines (RBTL)
A method used to extract a single representative spectrum at maximum light by ignoring strong absorption lines. This step is designed to be less prone to astrophysical bias compared to purely photometric methods.
Adaptive Spectroscopic Decontamination Module
A proposed refinement using a deep learning model that predicts and corrects for residual host galaxy contamination based on local metrics like dDLR or stellar mass M host, moving correction from fixed subtraction to dynamic learning.
Dynamic Weighting
Integrating the intrinsic spectral scatter function as a weighting mask in the DTEM process. This allows the system to automatically deweight wavelength bins with poor feature stability during model spectrum calculation.

Terminology

Summary

This paper presents the first application of the Twins Embedding (TE) spectrophotometric standardisation method to a large, heterogeneous dataset, addressing the challenge of using low-resolution spectra for precise cosmological measurements. By utilizing 3069 spectra from the Zwicky Transient Facility’s (ZTF) SEDm Integral Field Spectrograph (IFS), this study aims to build a large homogeneous spectrophotometric Type Ia supernova (SN Ia) sample and demonstrate the feasibility of achieving high-precision standardisation even when dealing with limited spectral quality. The results confirm that spectroscopic methods can yield low Hubble residual scatter, offering an alternative to photometric techniques that are less prone to astrophysical biases.

Data Preparation and Flux Calibration

The study begins by addressing the limitations of the SEDm instrument, which is not designed as a spectrophotometric instrument. To overcome this lack of intrinsic flux calibration, the researchers leverage high-quality ZTF photometric data. The process involves several critical steps to ensure accuracy:

  1. Interpolating lightcurve (LC) data using SALT2.4 to create a time-dependent model.

  2. Comparing the synthetic photometry calculated from the spectra in the ZTF g, r, i filters against these interpolated fluxes at the observed times.

  3. Correcting for second-order polynomials fitted through the three photometric points to achieve flux calibration accuracy at the percent level.

This rigorous process yields a final sample of 1897 flux-calibrated spectra from 1607 SNe Ia, with an estimated photometric accuracy of 0.07 mag in the 4500–7000 Å range.

The Read Between The Lines (RBTL) Method

The first step in standardisation involves applying the Read Between The Lines (RBTL) method to extract a single, representative spectrum at maximum light for each SN Ia. This is achieved by using the Differential Time Evolution Model (DTEM), which approximates the quadratic evolution in phase of SN Ia per wavelength with respect to maximum light within the plus or minus 5 days rest-frame phase range.

Once a single, corrected spectrum is RBTL-fitted, the color term (AV) and magnitude offset (m are estimated by minimizing a chi squared function. The model accounts for chromatic intrinsic dispersion (eta(lambda)) and the calibration uncertainty. This approach is designed to capture the color term ignoring the strong absorption lines, allowing researchers to isolate variability that is common across all SNe Ia, thereby ensuring that this method is less prone to astrophysical bias than photometric methods.

Twins Embedding (TE) Standardisation

The second phase of standardisation involves using non-linear parameters derived from the Twins Embedding (TE) technique. After the RBTL correction, the spectra are input into a Manifold Learning process using an Isomap algorithm. This technique is designed to capture variabilities in SN Ia spectra at maximum that are not already accounted for by the color and magnitude terms.

The final stage uses a Gaussian Process (GP) to predict the remaining gray offsets (delta mGP) based on these three Isomap parameters (xi 1, xi 2, xi 3). This allows for a complete correction of the spectrum:

mu TE,i = - m i - beta RBTL times AV i - delta mGP(xi i)

This method aims to reduce the scatter further, though the study notes that the residuals dispersion does not improve for the ZTF sample due to high instrumental error.

Results and Limitations

The initial RBTL standardisation on the ZTF sample of 783 SNe Ia resulted in a Hubble residual scatter of 0.153 mag (nMAD). When accounting for redshift error and spectrophotometric uncertainty, this value is estimated at an upper limit of about 0.129 mag. The TE method, while theoretically powerful, yielded a higher residual scatter for the ZTF sample than the SNfactory data (which achieved 0.097 mag).

The study also investigated host galaxy contamination by analyzing the correlation between residuals and host properties. The results showed that the amplitude of the astrophysical bias is strongly reduced when using RBTL standardisation, achieving steps that were consistent with zero, confirming the superiority of this spectroscopic approach over SALT photometric methods.

Improvements for AI systems

Based on the rigorous analysis of this manuscript, I have identified several critical areas where current AI systems can be significantly enhanced to address the specific challenges inherent in spectrophotometric standardisation (e.g., low resolution, host contamination, and feature scarcity).

The following improvements outline how an advanced AI system can move beyond simple data processing to achieve sophisticated astrophysical inference:

Improvement: Develop a specialized Deep Learning module that models the relationship between local environmental metrics (e.g., dDLR, stellar mass M host) and the required spectral correction coefficients (a 0, a 1). This moves beyond simple subtraction in Hypergal.

What the improved AI system can do: It will dynamically predict and correct for residual host galaxy contamination on a per-spectrum basis. Instead of relying on fixed cuts, it learns that specific a 1 slopes (which correlate with local color/reddening) are strongly associated with low dDLR values, allowing the system to proactively flag and correct SNe that are physically closer to their host galaxy before they introduce scatter into the final standardisation.

Improvement: Integrate a Dynamic Weighting layer into the Differential Time Evolution Model (DTEM). This layer utilizes the intrinsic spectral scatter function, eta(lambda), as a dynamic weighting mask.

What the improved AI system can do: When performing phase correction, it will not treat all wavelengths equally. If eta(lambda) indicates high chromatic dispersion (i.e., strong absorption lines or features poorly constrained by the low-resolution SEDm), the AI automatically deweights those unstable wavelength bins during the calculation of f model, i(lambda). This ensures that the resulting spectrum at maximum light is defined by its stable spectral domains, leading to a more robust and less artifact-prone RBTL parameter estimation (AV).

Improvement: Implement a transfer learning mechanism for the Isomap manifold learning process. Instead of relying solely on pre-trained vectors from the SNfactory dataset, this system will adapt the Isomap structure based on the observed spectral quality and feature set (e.g., presence/absence of Ca II lines) of the target dataset (ZTF).

What the improved AI system can do: It will identify when a new dataset lacks critical features used in training (like those missing in ZTF's specific wavelength range). It then adaptively adjusts the manifold parameters (xi 1, xi 2, xi 3) to account for this feature scarcity. This prevents the lack of information seen in Section 6.2, allowing the AI to achieve a lower residual dispersion even when working with lower spectral quality data than the original SNfactory sample.

Improvement: A comprehensive uncertainty quantification module that calculates a total error floor (sigma total) as a quadrature sum of three distinct, measurable components:

  1. Flux Calibration Floor (sigma calib): Derived from the residual a 0, a 1 dispersion (e.g., about 0.04 mag).

  2. SNR Variance (sigma SNR): Based on the observed binning of the spectrum (e.g., about 0.057 mag).

  3. Model Misfit (sigma model): The unexplained residual variance after applying RBTL/TE corrections.

What the improved AI system can do: It provides a dynamic, data-driven precision estimate for every single SN Ia. Instead of providing a fixed 0.07 mag error floor, it tells the user: This specific SN has an estimated uncertainty of 0.129 mag, which is driven by SNR limitations (65%) and flux calibration residual bias (35%). This allows for optimal data selection and provides a statistically rigorous measure of the reliability of the final distance estimate.

Abstract

This paper has two aims: the first one is to build a large homogeneous spectrophotometric Type Ia supernova (SN Ia) sample, using 3069 spectra from the second Zwicky Transient Facility data release (ZTF DR2). Using this sample we reproduce, as the second objective of the paper, the Twins Embedding (TE) spectrophotometric standardisation method, which led to an exceptionally low value of 0.073 mag for the intrinsic scatter. We improve the flux-calibration accuracy of the SEDm SN Ia spectral sample using the ZTF photometric data, which are calibrated at the percent level. We then apply the three steps of the TE parameterisation to a subset of 783 ZTF SN spectra near maximum light, and analyse the resulting standardisation methods. The precision of the phase correction model, which is the first step of the TE, is estimated at 0.01 mag in g band, using ZTF data. Despite the challenge posed by the ZTF spectrum extraction pipeline, we apply a first standardisation in color based on the second step of the TE, the Read Between The Lines (RBTL). When considering the scatter due to the redshift error and the flux calibration error, we estimate a 0.129 mag Hubble residual scatter for this ZTF sample as an upper limit. As expected from the low spectral quality, the final TE standardisation based on three non-linear parameters did not improve the overall dispersion. We release 1897 flux calibrated spectra of 1607 SNe Ia with an estimated photometric accuracy of 0.07 mag. We further demonstrate the ability to apply a spectrophotometric standardisation with limited quality spectra. The RBTL standardisation is more efficient than that of SALT with one less parameter, and the resulting host steps are consistent with zero, making it less prone to astrophysical bias. For future spectroscopic surveys, a better spectral quality would enable the full TE standardisation to be computed. (Abridged)

Sources

Related papers