Either squeeze more from brightness, or leave most data unused
For generations, astronomers have read the universe's expansion through the fading light of stellar explosions, yet the signals arriving from those distant supernovae have always carried more noise than clarity. A new artificial intelligence method called CIGaRS, developed by researchers at SISSA and the University of Barcelona, now attempts to untangle the overlapping physical stories written in a supernova's brightness alone — without the spectroscopic tools that have long been considered indispensable. As observatories prepare to flood the field with millions of new detections, this approach asks whether wisdom can be drawn not from the few carefully examined, but from the vast and imperfectly known.
- Upcoming surveys will detect over 100,000 Type Ia supernovae per year, yet fewer than one in a hundred will receive the detailed spectroscopic study traditionally needed to make sense of them.
- The light from each supernova carries tangled messages — dust, stellar age, chemical composition, cosmic distance — and separating them has historically demanded time and resources the field can no longer afford to spend.
- CIGaRS trains neural networks on massive simulated datasets, teaching the system to work backward from raw brightness measurements to the underlying physics of both the exploding star and its host galaxy simultaneously.
- Early tests show the method can sharpen cosmological measurements by a factor of four and may finally distinguish whether supernova brightness variations arise from dust, stellar age, or metallicity — a debate that has shadowed the field for decades.
- The researchers acknowledge that modeling assumptions about stellar evolution and dust laws could introduce bias, but argue that simulation-based inference itself offers a path to testing and containing those uncertainties.
For decades, Type Ia supernovae have served as the universe's mile markers — measure how bright a stellar explosion appears, compare it to how bright it should be, and distance reveals itself. Stack enough such measurements and the history of cosmic expansion comes into view. The difficulty has always been that a supernova's light carries multiple overlapping stories at once: the nature of the explosion, the age and chemistry of the star, the dust of its host galaxy, the stretching of space itself. Spectroscopy — breaking light into its component wavelengths — has long been the tool for sorting these signals apart.
Konstantin Karchev and Roberto Trotta at SISSA, together with Raúl Jiménez of the University of Barcelona, have built a different kind of instrument. Their method, CIGaRS, uses neural networks to model supernovae and their host galaxies together, disentangling intrinsic stellar properties from environmental effects using brightness data alone. The timing is deliberate. The Vera Rubin Observatory in Chile will soon begin detecting at least 100,000 Type Ia supernovae per year over a decade-long survey — yet only around one percent are expected to receive spectroscopic follow-up. Without a photometric-only approach, most of that incoming data would go scientifically unused.
Rather than treating dust, distance, galaxy evolution, and measurement noise as separate correction steps, CIGaRS folds them into a single forward model. Neural networks trained on vast simulated catalogs learn to reason backward from observed light to the physical and cosmological parameters underneath. Tested on a mock dataset of roughly 1,578 supernovae — comparable to today's flagship samples — the method simultaneously recovered cosmic expansion parameters, stellar delay-time distributions, and host-galaxy influences. Scaled to 16,000 objects, approximating a single month of Rubin data, it revealed something the standard mass-step correction had been quietly obscuring: metallicity and stellar age leave distinct fingerprints in brightness data, and lumping them together as a single correction blurs genuinely different physical causes.
The precision gains are striking. Cosmological constraints could sharpen by a factor of four compared to spectroscopic-only subsets, with two-dimensional parameter regions shrinking tenfold in larger samples. Photometric redshift estimates emerged as a byproduct of unusual quality. Trotta described the shift plainly: collecting detailed spectra for millions of objects will simply be impossible, and CIGaRS offers a way to reason through nearly the full sample instead of a narrow slice. The researchers are candid that modeling assumptions about stellar evolution and dust laws carry risks as datasets grow more precise — but they argue that simulation-based inference is itself a tool for testing those assumptions rather than hiding them. What the work ultimately proposes is that the coming era of astronomy will require not just more data, but fundamentally new ways of thinking through data that is vast, tangled, and imperfect.
For decades, astronomers have used Type Ia supernovae as cosmic mile markers. The method is straightforward in principle: measure how bright a stellar explosion appears from Earth, compare it to how bright it should be, and you can calculate the distance. Stack enough of these measurements together and you begin to see how the universe has expanded across billions of years. The trouble is that the light arriving from a supernova carries multiple overlapping messages at once, and teasing them apart has always required painstaking work.
The brightness of a supernova tells you about the explosion itself, but also about the star that exploded—its age, its chemical composition. The light gets dimmed and reddened by dust in the galaxy where the blast occurred. It gets stretched by the expansion of space itself. For years, astronomers have relied on spectroscopy—breaking light into its component wavelengths—to sort through these tangled signals. But a new approach, developed by Konstantin Karchev and Roberto Trotta at SISSA, along with Raúl Jiménez of the University of Barcelona, attempts to do something different. Their method, called CIGaRS, uses artificial intelligence and neural networks to model supernovae and their host galaxies together, pulling apart the intrinsic properties of the exploding star from the environmental factors surrounding it. The key innovation: it does all this using only brightness measurements, no spectroscopy required.
The timing of this breakthrough matters enormously. The Vera Rubin Observatory in Chile, which will begin operations soon, is expected to discover millions of supernova candidates over a decade. Each year alone, it should detect at least 100,000 Type Ia events. Yet the researchers estimate that only about 1 percent of these will receive spectroscopic follow-up observations—the detailed, time-consuming measurements that have traditionally been essential for understanding what you're looking at. Even with dedicated campaigns, only up to 10 percent of the host galaxies may be studied in detail. The field faces a stark choice: either extract far more information from brightness data alone, or watch most of the incoming data go unused.
CIGaRS works by folding multiple layers of physics into a single forward model rather than handling them as separate correction steps. The framework incorporates galaxy evolution, dust extinction, cosmological distance, measurement noise, sample selection, and the rate at which Type Ia supernovae appear over cosmic time. It also accounts for the delay-time distribution—how long it takes for a star system to produce a Type Ia supernova after the stars first form. Instead of using traditional statistical sampling methods, which the team describes as impractical at realistic scales, the approach relies on simulation-based inference. Neural networks are trained on vast sets of simulated examples, teaching the system to work backward from observed light to the underlying physical and cosmological parameters.
When the researchers tested their method on a simulated catalog of 1,578 supernovae—roughly the size of current flagship datasets—CIGaRS successfully reconstructed multiple intertwined quantities simultaneously: the cosmological parameters governing cosmic expansion, the delay-time distribution, and host-related influences like stellar age and metallicity. The team then scaled the test to a second mock dataset of roughly 16,000 objects, approximating what Rubin might collect in a single month. The results revealed something important about how different physical effects leave their fingerprints in the data. Metallicity—a star's chemical composition—tends to create a signature that looks like the familiar "mass step," a jump in brightness around a host galaxy mass of about 10 billion solar masses. Stellar age, by contrast, produces a more gradual, linear trend. This distinction matters because it suggests that the standard mass-step correction, long used as a practical fix, actually blurs together multiple distinct physical causes.
The gains in precision are substantial. Compared with methods that rely only on the small subset of supernovae observed spectroscopically, applying CIGaRS to large photometric-only samples could sharpen cosmological constraints by roughly a factor of four. In the larger mock catalog, one-dimensional credible intervals became about three times narrower, while two-dimensional parameter regions shrank by a factor of ten. The method also produced unusually strong photometric redshift estimates as a byproduct—measurements of how far away objects are based on their colors alone. In the 1,578-object test sample, the median error was about 0.01 in redshift, with no outliers exceeding an absolute error of 0.1, rivaling analyses that use much broader wavelength coverage.
Roberto Trotta noted the fundamental shift this represents. "Today, spectroscopy remains the richest tool for studying a supernova," he said, "but collecting detailed, homogeneous spectra at multiple epochs for very large samples will be impossible, given the sheer volume of data expected in the coming years." What CIGaRS offers is a way to use nearly the full sample rather than leaning on a narrow slice. That would improve measurements of cosmic expansion without waiting for costly follow-up campaigns that cannot keep pace with incoming data. The work may also help settle long-running debates about whether differences in supernova brightness stem mainly from dust, progenitor age, metallicity, or some remaining mass-related effect—questions that matter for both supernova physics and dark energy studies.
The researchers are careful about the limitations. Their current framework depends on constrained modeling assumptions that may not fully capture reality. Choices about the stellar initial mass function, stellar evolution tracks, and dust-extinction laws can bias results, especially as datasets become larger and more precise. But they argue that simulation-based inference offers a practical way to test those assumptions and fold uncertainties into the analysis. What emerges is a broader lesson: the coming era of astronomy will not just bring more objects to count. It will demand fundamentally new ways to reason through tangled, imperfect, high-volume data. Supernova cosmology is becoming a test case for that future.
Notable Quotes
Collecting detailed, homogeneous spectra at multiple epochs for very large samples will be impossible, given the sheer volume of data expected in the coming years.— Roberto Trotta, SISSA
We will have enormous numbers of supernovae, datasets so rich that they can no longer be analyzed using traditional methods. What we need are new methods capable of analyzing these large catalogues in a robust way.— Roberto Trotta, SISSA
The Hearth Conversation Another angle on the story
Why does it matter that we can do this without spectroscopy? Isn't spectroscopy more accurate?
Spectroscopy is richer in detail, yes. But there's a practical wall coming. Rubin will find 100,000 Type Ia supernovae per year. You cannot spectroscopically follow up 100,000 objects. You can maybe do 1,000. So you either develop a method that works on brightness alone, or you throw away 99 percent of your data.
But brightness alone seems like it should be noisier, less reliable.
It would be, if you treated each supernova in isolation. The trick is modeling the whole system at once—the supernova, its host galaxy, the dust, the age of the progenitor star. When you fold all those into a single neural network trained on millions of simulations, the network learns to disentangle signals that look hopelessly mixed.
What's the practical payoff? Better measurements of how fast the universe is expanding?
That's one. But there's something deeper. Astronomers have argued for decades about why supernovae in heavy galaxies look different from those in light galaxies. Is it dust? Age? Chemical composition? This method can actually separate those causes. That settles a real physics question.
And the constraints on cosmic expansion—you said they could improve by a factor of four?
In the simulations, yes. That's the difference between having data on a few hundred supernovae with spectra versus having data on tens of thousands with brightness alone. The volume overwhelms the noise.
Does this work on real data yet, or just simulations?
Right now, just simulations. The real test comes when Rubin starts observing. But the framework is built to handle the kinds of uncertainties and biases that real data will introduce.