🔭 astrophysics

Probing evolution of Long GRB properties through their cosmic formation history aided by Machine Learning predicted redshifts

This study utilizes machine learning to predict redshifts for long Gamma-Ray Bursts lacking spectroscopic data, creating an augmented sample that reveals current evolutionary models fail to fully explain GRB formation rates at both low and high redshifts, suggesting the need for more complex evolution terms or alternative progenitor origins.

Original authors: Dhruv S. Bal, Aditya Narendra, Maria Giovanna Dainotti, Nikita S. Khatiya, Aleksander L. Lenart, Dieter H. Hartmann

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Dhruv S. Bal, Aditya Narendra, Maria Giovanna Dainotti, Nikita S. Khatiya, Aleksander L. Lenart, Dieter H. Hartmann

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

The Big Picture: Finding the "Lost" Long-Range Phone Calls

Imagine the universe is a giant, dark city. Gamma-Ray Bursts (GRBs) are like incredibly bright, short-lived fireworks exploding in the sky. They are so bright that we can see them from the very edge of the universe, even from when the universe was a baby.

Scientists love these fireworks because they act as messengers. By studying them, we can learn how stars were born and died throughout history. However, there is a big problem: to understand a firework, you need to know exactly how far away it is. In astronomy, this distance is called "redshift."

Usually, to measure this distance, scientists need to catch the "afterglow" (the fading light) of the firework and analyze its colors. But, just like trying to catch a firefly in a storm, this is hard. About 80% of these fireworks fade too quickly or are too dim for our telescopes to catch the afterglow. This leaves us with a huge pile of data where we know a firework happened, but we have no idea how far away it was.

The New Tool: The "Redshift Predictor" (Machine Learning)

Since we can't measure the distance for most of these events, the authors of this paper used a clever trick: Machine Learning (ML).

Think of the fireworks that do have known distances as a training class. Scientists taught a computer to look at the characteristics of the bright explosion (the "prompt" phase) and the fading light (the "afterglow") of these known fireworks. The computer learned patterns, essentially saying, "When a firework looks like X and fades like Y, it is usually at distance Z."

Once the computer was trained, it was sent out to predict the distances for the 80% of fireworks that were previously "lost." This created a much larger, more complete map of the universe's history.

The Investigation: Are Fireworks Just Star Birth Markers?

The main goal of the study was to answer a simple question: Do these fireworks happen exactly when and where new stars are being born?

The Theory: Most long-lasting fireworks (Long GRBs) are caused by massive stars collapsing into black holes. So, scientists expect the rate of fireworks to perfectly match the rate of star formation.
The Test: The team took their new, larger map (which included the computer-predicted distances) and compared it to the known "Star Formation Rate" (how many stars are being born).

They tested three different scenarios to see which one fit the data:

No Evolution: The fireworks behave exactly the same way today as they did billions of years ago.
Beaming Evolution: The fireworks are like flashlights. Maybe in the past, the flashlights were pointed differently (beamed) than they are now.
General Evolution: Everything about the fireworks (brightness, direction, etc.) changes over time according to a specific rule.

The Results: A Partial Match

The results were a mix of success and mystery:

The Middle Ground: In the middle of the universe's history (redshift 1 to 2), the computer-predicted data worked well. The different scenarios could explain why the fireworks happened at that rate.
The Edges (Too Early and Too Late): The models failed to explain the data at the very beginning of the universe (high redshift) and the very recent past (low redshift).
- The Analogy: Imagine trying to fit a straight ruler against a curved road. The ruler fits the middle section perfectly, but it leaves gaps at the start and the end. This suggests that the "rules" for how these fireworks behave might be more complex than a simple straight line.

Why the Mismatch?

The authors offer a few guesses for why the models didn't fit the edges perfectly:

The Sample is Mixed: Just like a bag of mixed nuts might contain both peanuts and cashews, the sample of fireworks might contain different types of explosions. Some might be from collapsing stars, while others might be from stars crashing into each other (mergers). If you mix them up, the pattern gets messy.
The "Low-Redshift" Hump: At the recent end of the timeline, the data showed a small "bump" or excess of fireworks. This could be caused by a specific type of low-power explosion or perhaps the way the data was calculated, but the authors say they need more study to be sure.
The "High-Redshift" Mystery: At the very beginning of the universe, the data didn't match the star formation rate as well as expected. This might be because our telescopes are missing the faintest, most distant fireworks, or because the galaxies hosting them are very different from the ones we see today.

The Conclusion

This paper is a major step forward because it successfully used Machine Learning to fill in the missing pieces of the cosmic puzzle. By predicting the distances of the "lost" fireworks, the scientists created a much bigger dataset.

However, the study concludes that our current understanding isn't perfect yet. The simple rules we use to predict how fireworks evolve over time don't explain the entire history of the universe. It suggests that either the fireworks themselves change in complex ways, or we are looking at a mix of different types of cosmic explosions that we haven't fully sorted out.

To solve the rest of the puzzle, the authors suggest we need even more data from future telescopes and perhaps combine different types of light (optical and X-ray) to train our AI models even better.

Technical Summary: Probing Evolution of Long GRB Properties Through Their Cosmic Formation History Aided by Machine Learning Predicted Redshifts

Problem Statement
Gamma-ray Bursts (GRBs) serve as critical probes for cosmic star formation history (SFRH) extending back to the epoch of reionization. However, a significant limitation in current studies is that approximately 80% of GRBs lack measured redshifts ( $z$ ), as spectroscopic or photometric determination requires rapid follow-up of afterglow emission which is often unattainable. Existing methods to estimate redshifts using internal GRB correlations suffer from circularity arguments (assuming a cosmology to derive the correlation, then using it to find $z$ ) and systematic uncertainties. Consequently, datasets used to derive the Long GRB Rate Density (LGRB-RD) are small and potentially biased, hindering a clear understanding of GRB evolution and their relationship with the Star Formation Rate Density (SFRD).

Methodology
This study utilizes a sample of Long GRBs (LGRBs) with redshifts predicted by Machine Learning (ML) models, trained on prompt and X-ray afterglow parameters (specifically X-ray plateau correlations) without relying on cosmological assumptions. The analysis proceeds through the following steps:

Data Selection: The authors construct two datasets:
- Predicted Sample (PS): 228 LGRBs with X-ray plateaus but no observed redshifts, where $z$ was predicted using the ML model from Narendra et al. (2025) (N25). Outliers with high redshift or luminosity errors were removed.
- Combined Sample (CS): 446 LGRBs comprising the PS plus 218 LGRBs with known spectroscopic redshifts and X-ray plateaus (the "Observed Sample" or OS).
- Cuts: Both samples were filtered to remove GRBs below the instrument's luminosity threshold (determined via Kolmogorov-Smirnov tests) and GRB 060614 (a merger-origin long GRB).
Rate Density Calculation:
- The authors apply the Efron-Petrosian (EP) method to correct for selection effects and derive the cumulative rate density distribution ( $\sigma$ ).
- To account for potential evolution in GRB properties (e.g., luminosity), a correction factor $L' = L/(1+z)^k$ is applied. The parameter $k$ is determined by finding the value where the Kendall- $\tau$ statistic is zero, ensuring luminosity and redshift independence.
- The LGRB-RD ( $\rho$ ) is calculated by differentiating the cumulative distribution and scaling by factors accounting for instrument sensitivity, field of view, uptime, and the total number of detected GRBs.
Theoretical Comparison:
- A theoretical rate density ( $\rho_{theor}$ ) is constructed by scaling the SFRD (Madau & Dickinson 2014) by conversion factors (collapsar fraction, supernova types, beaming) and an evolutionary term $(1+z)^\delta$ .
- The study tests three evolutionary scenarios: no evolution ( $\delta=0$ ), beaming evolution ( $\delta=1.6$ ), and a general evolution term where $\delta$ varies between -4 and 4.

Key Results

Evolution Significance: The analysis reveals significant evidence for evolution in GRB properties. The best-fit $k$ values are $2.2^{+0.8}_{-0.7}$ for the Predicted Sample (PS) and $2.5^{+0.4}_{-0.4}$ for the Combined Sample (CS), corresponding to significances of $\sim3\sigma$ and $\sim6\sigma$ respectively, relative to the no-evolution case ( $k=0$ ).
Rate Density Trends:
- The rate densities for both samples peak in the redshift range of $z \sim 1-2$ .
- The PS peaks at a higher redshift than the CS, attributed to the ML model's bias toward the mean of the training distribution, which was subsequently corrected.
- A low-redshift ( $z < 1$ ) excess (uptick) is observed in the CS, which the authors attribute to potential contributions from mergers, Low Luminosity GRBs (LLGRBs), or fitting artifacts.
Evolutionary Models:
- No single evolutionary model (power law in $1+z$ ) successfully explains the rate density across the entire redshift range for either sample.
- While specific values of $\delta$ can match the low- $z$ or high- $z$ ends of the data, none capture the full distribution. For the CS, $\delta = -0.3$ fits the low- $z$ trend, while $\delta = -0.9$ fits the peak and high- $z$ end.
- The peak of the CS rate density occurs at $z \sim 1.6$ (lookback time $\sim 9.5$ Gyr), while the MD14 SFRD peaks at $z \sim 1.9$ (lookback time $\sim 10.1$ Gyr). The authors note this discrepancy is within error margins but suggests potential influences from binary interactions extending progenitor lifespans.

Significance and Claims
The paper claims that the integration of ML-predicted redshifts is a "crucial step forward" in expanding the dataset available for studying GRB evolution. By augmenting the sample size, the authors demonstrate that:

Sample Homogeneity Issues: The inability of simple power-law evolution terms to explain the rate density across all redshifts suggests either a more complex evolutionary term or sample non-homogeneity (e.g., a mix of collapsar and merger-origin GRBs within the long-GRB classification).
SFRD Tracing: While LGRBs generally trace the SFRD up to $z \sim 6$ , the specific evolution of their properties (beaming, metallicity, progenitor type) introduces deviations that simple scaling cannot fully resolve.
Future Directions: The authors emphasize that the current ML analysis is the most accurate to date but that future improvements require combining optical and X-ray plateau samples for training and utilizing data from upcoming missions (SVOM, Rubin, THESEUS, Hi-Z GUNDAM) to increase the sample of known redshifts. They caution that combining multi-wavelength samples requires care due to potential differences in emission sites and dust extinction effects.

The study concludes that while the derived rate densities are reliable in the mid-redshift range, the mismatches at low and high redshifts highlight the need for further investigation into progenitor diversity and the physical origins of the observed rate density features.