astronomy/ai produced the result/Monthly Notices of the Royal Astronomical Society 2024 · v2
Software reconstructs supernova spectra and physical properties from brightness measurements alone
Astronomers built CASTOR, an open-source tool that turns multi-band brightness measurements of an exploding star into a time series of synthetic spectra and a set of physical parameters. Gaussian Process regression does the filling-in.
spectrum · one line per step, placed by what the step does · bright lines used AI
Building spectral templates and reconstructing parameters for core-collapse supernovae with CASTOR
Monthly Notices of the Royal Astronomical Society, 2024
doi:10.1093/mnras/stae1911 · record aix-00048 v2 · checked 2026-10-08
- AI was for
- Denoising
- Model family
- Gaussian process
- Checked by
- Replication1 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
When a massive star runs out of fuel, its core collapses and the star explodes. Astronomers call this a core-collapse supernova. To work out what sort of star it was and how violent the explosion was, they want two kinds of measurement. One is photometry: how bright the event looks through a set of coloured filters, night after night, which gives a light curve. The other is spectroscopy: the brightness spread out wavelength by wavelength, which reveals the chemical elements present and how fast the debris is moving. Photometry is comparatively cheap and plentiful. Spectra need more telescope time, and for many supernovae only a few are taken, or none.
The researchers set out to bridge that gap with a piece of software called CASTOR. The idea is to take a supernova for which only multi-band photometry exists, and produce from it both a run of synthetic spectra across time and a list of physical quantities: the time of the explosion, how much dust lies in the way, the expansion speed, the distance, the luminosity, the mass of ejected material and of radioactive nickel, the temperature and size of the glowing surface, and the radius and mass of the star that blew up.
Where AI came in
The learned part of the pipeline is Gaussian Process regression, a statistical method that draws a smooth curve through scattered measurements and also reports how uncertain it is at each point. It is used twice. First in one dimension, to turn the dotted light curves of each supernova into continuous ones in every filter. Those smoothed curves let the software compare a new supernova against a catalogue of 111 past ones and pick the closest match as a reference object. Here it stands in for fitting a fixed mathematical formula, such as a polynomial, to the light curve.
Then in two dimensions, across time and wavelength at once. The software lays out a grid combining the new supernova's calibrated brightnesses with some of the reference supernova's real spectral points, and the Gaussian Process interpolates across the gaps to produce the synthetic spectra. In this step it stands in for spectroscopic follow-up observation itself. Everything after that is conventional physics and curve fitting rather than learning. The authors ran the tool on the type Ib supernova SN2015ap, for which SN2016iae was chosen as the reference and 50 synthetic spectra were built, and compared the resulting parameters with published values for the same object.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
CASTOR is an open-source data-analysis tool that reconstructs synthetic spectra and physical parameters of core collapse supernovae from multi-band photometry alone. It uses Gaussian Process regression to interpolate light curves and, after selecting the photometrically most similar supernova from a catalogue of 111 core collapse supernovae by a normalised chi-squared test, interpolates photometric fluxes together with that reference object's spectra in time and wavelength to build a spectral time series. Physical parameters of the event, the ejecta, the photosphere and the progenitor are then derived from the interpolated light curves and the synthetic spectra using empirical relations and standard physical assumptions. The authors demonstrate the software on the type Ib supernova SN2015ap, for which SN2016iae was selected as reference and 50 synthetic spectra were built, and compare the resulting parameters with values published for the same object.
How AI was used
Gaussian Process regression with Matérn (ν=1.5) kernels is used in two places. First, one-dimensional GPs, fitted with the george package, interpolate the multi-band light curves of both the training-set supernovae and the newly observed supernova, using composite kernels whose lengthscales are set from the medium, maximum and minimum spacing between consecutive photometric points; the interpolated training-set curves are resampled at the observed epochs so that a normalised chi-squared can be computed filter by filter to pick a reference supernova. Second, a two-dimensional GP implemented with scikit-learn interpolates, across time and wavelength, a grid combining the calibrated flux densities derived from the interpolated light curves of the case-of-study supernova with part of the spectral points of the reference supernova, producing the synthetic spectral time series; wavelength lengthscale is fixed at 70 Å and time lengthscales are taken from the minimum and maximum spectral sampling steps. Magnitudes are calibrated to spectrometric magnitudes beforehand using a slope and intercept obtained by a linear fit over the whole training set. The subsequent parameter estimation uses non-learned procedures — polynomial fits to P-Cygni profiles with user-selected line intervals, Cardelli's law, Hubble's law, a linear fit of bolometric luminosity against the energy deposit function, and a dilution-corrected black-body fit to the spectral energy distribution.
The shape of the work
Structural · the record, drawn
no AI
Assemble CCSN training catalogue
Obtaining raw data, whether by measurement, download or retrieval.
CASTOR employees a catalogue of 111 CCSNe to select the reference supernova from.where the paper describes this · verbatim
AI
Interpolate multi-band light curves with Gaussian Processes
Running a trained model over new data to predict, classify or score. The AI stood in for statistical model.
we interpolate the multi-band photometry of the training set using the GPswhere the paper describes this · verbatim
no AI
Select reference supernova by chi-squared comparison
Reducing a candidate set by filtering or ranking, in a single pass.
Finally, the supernova with the lowest normalized chi-squared is identified as the reference supernova.where the paper describes this · verbatim
no AI
Derive photometry-to-spectroscopy calibration parameters
Cleaning, filtering, normalising or labelling data already obtained.
We designed a calibration technique specifically for converting photometric magnitudes into spectrometric magnitudes and vice-versa.where the paper describes this · verbatim
AI
Build synthetic spectral templates with two-dimensional GP regression
Producing candidate objects that did not previously exist. The AI stood in for physical experiment.
we interpolate flux points in the time and wavelength domains, by means of two-dimensional GP regression method on the flux gridwhere the paper describes this · verbatim
no AI
Estimate event, ejecta and progenitor parameters
Extracting understanding from model behaviour.
CASTOR employees only the output described in Section 2: the interpolated light curves and the spectral templateswhere the paper describes this · verbatim
no AI
Compare reconstruction for SN2015ap with published values and observed spectra
Testing outputs against ground truth.
studying available data from SN2015ap and comparing our results with those available in literaturewhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
CASTOR builds synthetic spectral templates by means of Gaussian Process regression techniqueswhere the paper describes this · verbatim
The synthetic spectral templates that the paper is about are produced by Gaussian Process regression; the downstream parameter map is derived from those templates.
CASTOR combines Gaussian Process and other Machine Learning techniques to build time-series templates of synthetic spectrawhere the paper describes this · verbatim
We compare the results of our test with those from.where the paper describes this · verbatim
The source code and the observational data compiled for this project are publicly available at the online repository.where the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- Version of Gaussian Process regression with Matérn (ν=1.5) kernel, george implementation (light curves)Which version of the model was used is not stated.
- Version of Gaussian Process regression with Matérn (ν=1.5) composite kernel, scikit-learn implementation (spectra)Which version of the model was used is not stated.
About this article
Record aix-00048, version 2, checked by a person on 2026-10-08. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error