Revisiting the Hubble Tension with DESI DR2 Baryon Acoustic Oscillation Observations and Machine Learning Methods

arXiv:2610.08836 · astro-ph.CO · Submitted 2026-09-29 · 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: We're looking at the paper: "Revisiting the Hubble Tension with DESI DR2 Baryon Acoustic Oscillation Observations and Machine Learning Methods".

Jocelyn: Here's the gist of the paper. One of the key unresolved questions in modern cosmology is the Hubble tension, which arises from the inconsistency between determinations of the Hubble constant (H0) based on nearby observations and those predicted by early-Universe data under the standard ΛCDM paradigm Our results exhibit excellent agreement with the recent Planck measurements within 1σ, which highlight the potential of data-driven, model-independent methods to address the Hubble tension The paper utilizes the latest DESI DR2 baryon acoustic oscillation (BAO) measurements combined with nonparametrically reconstructed Type Ia supernovae and cosmic chronometer data to constrain the Hubble constant using three complementary machine learning methodologies: Gaussian process regression (GPR), artificial neural networks (ANNs), and long short-term memory (LSTM) networks.

Vera: The Hubble constant is estimated by combining the reconstructed cosmic chronometer function H(z), the luminosity distance from SN Ia, and the angular diameter distance from BAO measurements using the distance duality relation (DDR).

Jocelyn: This formulation allows for an independent estimation of H0 in terms of three independent observables: H0 = 1 /

(1 + z)2 × ΞSNIa(z): /

H(z)dA(z): .

Vera: Machine Learning Reconstruction Techniques The study employs several nonparametric reconstruction techniques to handle the data and reconstruct the expansion history H(z) from the cosmic chronometer (CC) dataset.

Title and authors: Vera: So we're looking at this paper called "Revisiting the Hubble Tension with DESI DR2 Baryon Acoustic Oscillation Observations and Machine Learning Methods". It’s trying to tackle that big problem where measurements of how fast the universe is expanding near us don't match what early-universe data predicts.

Jocelyn: Yeah, it seems they are using this new DESI data combined with some machine learning tools to see if they can find a way around that mismatch. It’s about finding a more model-independent way to figure out the Hubble constant H0.

Subrahmanyan: That tension is huge because it challenges the standard ΛCDM model, which suggests early dark energy or phantom energy might be needed to fix it, but those extensions have their own problems when you look at other data.

Vera: Exactly. This paper focuses on using different types of measurements together—the cosmic chronometer function, supernova luminosity distances, and the baryon acoustic oscillation data from DESI DR2—to try and get an independent estimate of H0.

Jocelyn: And they’re doing this by combining these inputs into a distance duality relation framework to estimate H0 using three different measurement types simultaneously. It’s a clever way to cross-check things that are usually done separately.

Subrahmanyan: The core idea here is to test the redshift consistency of H0 and also look for potential systematic biases across different cosmic epochs, which is something you need when you are trying to disentangle observational errors from actual cosmological anomalies.

Vera: They use three machine learning techniques—Gaussian process regression, artificial neural networks, and long short-term memory networks—to reconstruct the expansion history function H(z) from the cosmic chronometer data. That’s where the non-parametric part comes in.

Jocelyn: The paper talks about how these different AI methods are optimized for different kinds of data; for example, they might choose a simpler architecture if the cosmic chronometer data is sparse, or a more complex one if they have lots of supernova measurements.

Subrahmanyan: The goal with these machine learning tools is to get that smooth expansion history function H(z) without having to assume any specific background cosmology beforehand; it’s about mapping redshift directly to the Hubble parameter.

Vera: Then, they perform a cross-method comparison of the results. They find three different values for H0 based on GPR, ANNs, and LSTMs at the effective redshifts of DESI DR2 tracers.

Jocelyn: And what’s interesting is that when you look at how these different methods constrain H0 together by systematically excluding individual DESI BAO tracers, the central values only shift by less than two point five km s minus one Mpc per method.

Subrahmanyan: So, the joint results are pretty robust because they aren't overly dependent on any single measurement; that’s a good sign when you’re trying to solve a problem this fundamental to cosmology.

Vera: It suggests that these data-driven, model-independent methods have some potential for addressing the tension by providing better constraints than relying on just one type of measurement.

Title and authors: Jocelyn: It opens up a path where we can integrate these nonparametric reconstruction frameworks into AI systems to draw constraints from different data sets like DESI BAO and supernovae.

Subrahmanyan: And they’re using a Markov chain Monte Carlo strategy to sample the reconstructed distributions of the cosmic chronometer function, supernova data, and BAO measurements, which avoids having to run a full grid-like Cartesian combination of all these things.

Vera: They are also optimizing these systems by minimizing a risk metric defined as Bias squared plus Variance, which is designed to balance how well the reconstruction reproduces the unknown function against its smoothness across different test cases.

Jocelyn: It seems like a way to rigorously evaluate if the reconstruction actually captures the true expansion history and not just some noise in the data.

Subrahmanyan: This paper shows that even when you are looking at these complex, multi-method approaches, you can still achieve subpercent statistical precision when testing the redshift consistency of H0 and probing potential systematics across cosmic epochs.

Vera: The main implication for us is that this synthesis offers a new pathway to disentangle observational biases from actual cosmological anomalies using these diverse probes.

Jocelyn: It suggests that even if we can't immediately solve the Hubble tension, these methods give us a more rigorous test of the consistency of different ways we measure cosmic expansion.

Subrahmanyan: Ultimately, this work on "Revisiting the Hubble Tension with DESI DR2 Baryon Acoustic Oscillation Observations and Machine Learning Methods" shows that there are new pathways to constrain cosmology by combining data from different tracers like DESI DR2 BAO with SN Ia and cosmic chronometer data.

Vera: It’s a testament to how these AI tools can be used for joint constraint analysis without relying on a single, overly simplistic model.

Jocelyn: So we see here that the derived H0 central values are sixty-one point eight plus or minus five point three km s minus one Mpc per Gaussian process regression, while artificial neural networks give us sixty-six point three plus or minus five point two and LSTM networks give us a slightly higher value of sixty-eight point seven plus or minus some six point zero.

Subrahmanyan: And the paper confirms that these different estimates are consistent enough that they don't seem unduly biased by any single measurement when you look at the joint constraints on H0 across all methods.

Vera: It really highlights the potential of data-driven, model-independent methods to help us navigate this tension, even if we aren't solving it completely in this study.

Jocelyn: We’re going to keep an eye on how these specific machine learning architectures perform when applied to other cosmological datasets that might offer different strengths or weaknesses.

Subrahmanyan: It gives us a concrete set of tools to test theoretical extensions, so if we see a persistent deviation in the future, we know exactly where the systematic uncertainty might be coming from.

The paper's summary: Vera: So, to wrap up what we've just talked about, this paper is basically taking some new data from DESI DR2, which are these measurements of how galaxies are spaced out—BAO—and mixing that with other cosmic probes like supernova distance measurements and the cosmic chronometer function to try and get a better handle on how fast the universe is expanding.

Jocelyn: Right, so they're using this combination of data to create a framework where they can estimate the Hubble constant, H zero, in three different ways all at once by combining those distances together.

Subrahmanyan: The main point here is that they use three distinct machine learning methods—Gaussian process regression, artificial neural networks, and long short term memory networks—to try and reconstruct the expansion history of the universe without assuming any specific model beforehand.

Vera: That’s a big deal because it means they aren't stuck with one way of modeling things; they are letting the data guide them to find that smooth expansion history function, H zero as you know it.

Jocelyn: And what they did next was compare the results from these three different AI methods, and their findings for H zero were pretty close—they landed around sixty-one point eight km per second per megaparsec in one case, sixty-six point three in another, and a bit higher at some places.

Subrahmanyan: But they also looked at the joint constraints; when you look at how these different methods constrain H zero together by systematically excluding individual DESI BAO tracers, the central values only shift by less than two point five km per second per megaparsec across all those methods.

Vera: That tells us that the individual measurements aren't pulling in wildly different directions; they’re actually quite consistent when you look at them together.

Jocelyn: It shows that this data-driven approach, using these machine learning reconstructions, offers a way to get a more robust constraint on H zero than just relying on one specific type of measurement.

Subrahmanyan: I think what this means for the bigger picture is that it gives us a strong test of whether the different ways we measure cosmic expansion are actually consistent with each other across different scales and epochs.

Vera: It’s not about solving the tension right now, but it’s definitely giving us a better way to check if our fundamental measurements are actually sound before we start chasing those complex new dark energy theories.

Jocelyn: And they also talked about how these AI systems can be optimized for different kinds of data; they could use a simpler AI setup if the cosmic chronometer data is sparse, or something more complex for massive supernova datasets.

Subrahmanyan: That’s important because it shows that these methods are flexible; they aren't just one-size-fits-all, which is exactly what you need when you’re dealing with messy observational data from different sources.

Vera: So, the paper is setting up a framework where we can integrate these reconstruction techniques into AI systems to draw constraints from different data sets like DESI BAO and supernovae.

Jocelyn: And they're using a Markov chain Monte Carlo strategy to sample the reconstructed distributions of those three different measurements—the cosmic chronometer function, the supernova data, and the BAO measurements—which avoids having to run that full grid-like Cartesian combination of everything.

Subrahmanyan: That sampling approach is key because it lets us look at how these inputs interact without getting bogged down in a massive amount of computation.

Vera: It’s a solid piece of work showing that combining different data tracers with AI can give us constraints on cosmology that are really independent, which is what we need to tackle this tension head-on.

Jocelyn: So, moving forward, we need to see how these specific machine learning architectures perform when applied to other cosmological datasets that might offer different strengths or weaknesses.

The paper's improvements: Tom: So, we’ve looked at how this paper uses machine learning to get those estimates for H zero from different data sources—the BAO, the supernovae, and the cosmic chronometer—and now we're going to look at what they suggest we should do next with this technique.

Vera: The authors are suggesting a few ways to improve this whole process, mostly focusing on how these AI systems can be used more generally for cosmology.

Jocelyn: They think we should integrate these nonparametric reconstruction frameworks directly into the AI systems so they can derive model-independent constraints on cosmological parameters by combining data from different tracers like DESI DR2 BAO with supernova and cosmic chronometer data.

Subrahmanyan: That’s a big step because it means we move away from just using one specific measurement type to find the expansion history, H zero as you know it.

Vera: And they also propose a smarter way for the AI system to handle different datasets, saying that it should be optimized for different kinds of data by choosing an architecture whose shape fits the density and covariance complexity of those inputs.

Jocelyn: They even suggest using a single layer architecture with four thousand ninety-six neurons if the cosmic chronometer data is sparse, but maybe a dual layer with four thousand ninety-six neurons per layer if you have massive supernova datasets.

Subrahmanyan: That shows they’re thinking about how to make these models practical for real use on different observational setups.

Vera: And they are also looking at how the AI can actually check its own work, by minimizing a risk metric defined as Bias squared plus Variance, which balances how well the reconstruction reproduces the unknown function against its smoothness across different test cases.

Jocelyn: That’s basically a way to evaluate if the reconstruction is actually capturing the true expansion history and not just some noise that happened to look good on paper.

Subrahmanyan: It’s about making sure that when we use these AI tools, we aren't just getting an answer that looks right for one set of data but fails when you test it against something different.

Vera: So, the implication is that these methods could become a standard toolkit for combining heterogeneous observational data to get more reliable constraints on fundamental cosmological parameters.

Jocelyn: And they are using a Markov chain Monte Carlo strategy to draw samples from the reconstructed multivariate distributions of those three inputs—the cosmic chronometer function, supernova data, and BAO measurements—which avoids having to run that full grid-like Cartesian combination of everything.

Subrahmanyan: That sampling approach is key because it lets us look at how these inputs interact without getting bogged down in a massive amount of computation.

Vera: It’s a solid piece of work showing that combining different data tracers with AI can give us constraints on cosmology that are really independent, which is what we need to tackle this tension head-on.

Jocelyn: So, moving forward, we need to see how these specific machine learning architectures perform when applied to other cosmological datasets that might offer different strengths or weaknesses.

Conclusion: Tom: So, to wrap up this whole session on "Revisiting the Hubble Tension with DESI DR2 Baryon Acoustic Oscillation Observations and Machine Learning Methods," we've seen how combining those different observational probes with machine learning can give us a more robust handle on how fast the universe is expanding near us.

Vera: It really shows that these data-driven, model-independent methods have some real potential for testing the consistency of different ways we measure cosmic expansion.

Jocelyn: This study gives us a concrete set of tools to test theoretical extensions, so if we see a persistent deviation in the future, we know exactly where the systematic uncertainty might be coming from.

Subrahmanyan: I think what this means is that even if we can't solve the tension right now, this method helps us rigorously check if our fundamental measurements are actually consistent across different cosmic epochs.

Vera: Exactly. It’s a testament to how these AI tools can be used for joint constraint analysis without relying on just one type of measurement.

Jocelyn: We’re going to keep an eye on how these specific machine learning architectures perform when applied to other cosmological datasets that might offer different strengths or weaknesses.

Subrahmanyan: And this work lays a groundwork for testing those theoretical models we've been talking about, so it’s useful context for what comes next in the field.

Chenfa Zheng, Shuo Cao, Wuzheng Guo, Qiumin Wang, Xinyue Jiang, Marek Biesiada, Tonghua Liu

Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University · School of Physics and Astronomy, Beijing Normal University · National Centre for Nuclear Research, Pasteura

astro-ph.CO

Submitted: 2026-09-29

Updated: 2026-09-29

Comments: 17 pages, 9 figures, 4 tables. Published in The Astrophysical Journal Supplement Series

Journal ref: The Astrophysical Journal Supplement Series, 285:63 (2026)

DOI: 10.3847/1538-4365/ae7d1a

Code: https://github.com/dfm/emcee

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

Importance score: 78/100

The gist: One of the key unresolved questions in modern cosmology is the Hubble tension, which arises from the inconsistency between determinations of the Hubble constant (H0) based on nearby observations and…

Key concepts

Hubble Tension
This is a major problem in cosmology where measurements of the universe's expansion rate (H0) from nearby objects disagree significantly with predictions based on early-universe data under the standard cosmological model. The paper investigates if new, data-driven methods can reconcile these conflicting values.
Baryon Acoustic Oscillations (BAO)
BAO are characteristic patterns in the distribution of matter in the universe, imprinted by sound waves in the early cosmos. These patterns act as a 'standard ruler' that allows astronomers to measure distances and constrain cosmological parameters like H0 when combined with other distance measurements.
Gaussian Process Regression (GPR)
GPR is a machine learning technique used here to reconstruct the smooth expansion history of the universe, H(z), from cosmic chronometer data. It uses a kernel function to model the underlying function's smoothness and provides a strictly model-independent way to infer cosmological parameters.
Distance Duality Relation (DDR)
The DDR is a mathematical relationship used to combine different distance measurements—like luminosity distance from supernovae, angular diameter distance from BAO, and the expansion history H(z)—to independently estimate the Hubble constant. It allows for an estimation of H0 using three independent observables.

Terminology

Summary

One of the key unresolved questions in modern cosmology is the Hubble tension, which arises from the inconsistency between determinations of the Hubble constant (H0) based on nearby observations and those predicted by early-Universe data under the standard ΛCDM paradigm Our results exhibit excellent agreement with the recent Planck measurements within 1σ, which highlight the potential of data-driven, model-independent methods to address the Hubble tension

Methodology and Observational Framework

The paper utilizes the latest DESI DR2 baryon acoustic oscillation (BAO) measurements combined with nonparametrically reconstructed Type Ia supernovae and cosmic chronometer data to constrain the Hubble constant using three complementary machine learning methodologies: Gaussian process regression (GPR), artificial neural networks (ANNs), and long short-term memory (LSTM) networks. The Hubble constant is estimated by combining the reconstructed cosmic chronometer function H(z), the luminosity distance from SN Ia, and the angular diameter distance from BAO measurements using the distance duality relation (DDR). p. 2

Where the paper says it

METHOD AND OBSERVATIONS The Hubble constant can be independently estimated by combining the reconstructed cosmic chronometer function H(z), the luminosity distance from SN Ia, and the angular diameter distance from BAO measurements. Relying on the DDR, dL(z) (1 + z) 2dA(z) = 1, (1) which holds under the assumptions of photon number conservation and null geodesic propagation, we can formulate the Hubble constant in terms of three independent observables: H0 = 1 (1 + z) 2 ΞSNIa(z) ΞBAO(z) HCC(z), (2) where ΞSNIa(z) = H0dL(z) is derived from SN Ia apparent magnitudes (unanchored distance), ΞBAO(z) = H(z)dA(z) is obtained from the DESI DR2 BAO measurements, and HCC(z) is reconstructed at the corresponding redshift using CC data.

This formulation allows for an independent estimation of H0 in terms of three independent observables: H0 = 1 / [(1 + z)2 × ΞSNIa(z)] / [H(z)dA(z)]. p. 2
Where the paper says it

METHOD AND OBSERVATIONS The Hubble constant can be independently estimated by combining the reconstructed cosmic chronometer function H(z), the luminosity distance from SN Ia, and the angular diameter distance from BAO measurements. Relying on the DDR, dL(z) (1 + z) 2dA(z) = 1, (1) which holds under the assumptions of photon number conservation and null geodesic propagation, we can formulate the Hubble constant in terms of three independent observables: H0 = 1 (1 + z) 2 ΞSNIa(z) ΞBAO(z) HCC(z), (2) where ΞSNIa(z) = H0dL(z) is derived from SN Ia apparent magnitudes (unanchored distance), ΞBAO(z) = H(z)dA(z) is obtained from the DESI DR2 BAO measurements, and HCC(z) is reconstructed at the corresponding redshift using CC data.

Machine Learning Reconstruction Techniques

The study employs several nonparametric reconstruction techniques to handle the data and reconstruct the expansion history H(z) from the cosmic chronometer (CC) dataset. Specifically, Gaussian process regression (GPR) is implemented using a squared exponential kernel, employing a full-marginalization approach via Markov Chain Monte Carlo (MCMC) sampling to ensure a strictly model-independent reconstruction. Artificial neural networks (ANNs) are constructed as fully connected feed-forward networks, minimizing the L1 loss function (mean absolute error) after normalizing the input redshift and target data using z-score scaling. p. 4

Where the paper says it

The ANN is constructed as a fully connected feed-forward network. The sole input of the network is the redshift z, and the output is a two-dimensional vector containing the reconstructed theoretical function (e.

Long short-term memory (LSTM) networks are adapted to output the reconstructed function mean and its corresponding uncertainty by optimizing a Gaussian Negative LogLikelihood (NLL) loss with log-space variance supervision. p. 9
Where the paper says it

The training objective is completely shifted from a traditional parameter-fitting metric to a Gaussian Negative LogLikelihood (NLL) loss. To effectively balance the gradient weights between the highly precise low-redshift observations and the heavily uncertain high-redshift data, an explicit log-space variance supervision is incorporated into the objective.

Comparative Analysis and Constraints

The paper presents a cross-method comparative analysis of the derived Hubble constant, yielding H0 = 61.8 ± 5.3 km s−1 Mpc−1 (GPR), H0 = 66. p. 0

Where the paper says it

We perform individual and joint constraints on H0 at the effective redshifts of DESI DR2 tracers, yielding H0 = 61. 3 km s−1 Mpc−1 (GPR), 66.

3 ± 5.2 km s−1 Mpc−1 (ANN), and H0 = 68. p. 0
Where the paper says it

3 km s−1 Mpc−1 (GPR), 66. 2 km s−1 Mpc−1 (ANN), and 68.

7 ± 6.0 km s−1 Mpc−1 (LSTM) at the effective redshifts of DESI DR2 tracers. p. 0
Where the paper says it

In this paper, we use the latest DESI DR2 baryon acoustic oscillation (BAO) measurements, combined with nonparametrically reconstructed Type Ia supernovae and cosmic chronometer data, to constrain the Hubble constant using three complementary machine learning methodologies: Gaussian process regression (GPR), artificial neural networks (ANNs), and long short-term memory (LSTM) networks. We perform individual and joint constraints on H0 at the effective redshifts of DESI DR2 tracers, yielding H0 = 61.

The joint constraints on H0 by systematically excluding individual DESI BAO tracers show that the inferred H0 central values shift by merely ≲ 2. p. 0
Where the paper says it

In this paper, we use the latest DESI DR2 baryon acoustic oscillation (BAO) measurements, combined with nonparametrically reconstructed Type Ia supernovae and cosmic chronometer data, to constrain the Hubble constant using three complementary machine learning methodologies: Gaussian process regression (GPR), artificial neural networks (ANNs), and long short-term memory (LSTM) networks. We perform individual and joint constraints on H0 at the effective redshifts of DESI DR2 tracers, yielding H0 = 61.

5 km s−1 Mpc−1 across different methods, confirming that the joint results are not unduly biased by any single measurement. p. 8
Where the paper says it

The ANN reconstruction results applied to the mock CC datasets across four different network architectures. The single-layer configurations (panels (a) and (b)) demonstrate stable recovery of the fiducial cosmology, whereas the dual-layer configurations (panels (c) and (d)) exhibit severe overfitting behaviors, tightly winding around the noisy scattered points.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Nonparametric reconstruction frameworks can be integrated into AI systems to derive model-independent constraints on cosmological parameters by combining data from different tracers like DESI DR2 BAO with SN Ia and cosmic chronometer data.

  2. The AI system can perform joint constraint analysis by employing a Markov Chain Monte Carlo strategy that draws samples from the reconstructed multivariate distributions of HCC(z), SN Ia, and BAO measurements, avoiding the unnecessary computational burden of a full grid-like Cartesian combination.

  3. The system can leverage machine learning methods like Gaussian process regression (GPR), artificial neural networks (ANNs), and long short-term memory networks (LSTM) to reconstruct the smooth expansion history H(z) from CC data, providing a direct mapping from redshift to the Hubble parameter without assuming any background cosmology.

  4. The system can be optimized for different cosmological datasets by using architectures whose optimal topology is contingent upon the dataset's density and covariance complexity, such as selecting a single-layer architecture with 4096 neurons for sparse CC data versus a dual-layer architecture with 4096 neurons per layer for massive SN Ia datasets.

  5. The AI system can evaluate the truth-recovery capability by minimizing a risk metric defined as Risk = Bias2 + Variance, which balances the ability of the reconstruction to reproduce the unknown function f and its smoothness across different cosmological test cases.

Abstract

One of the key unresolved questions in modern cosmology is the Hubble tension, which arises from the inconsistency between determinations of the Hubble constant (H 0) based on nearby observations and those predicted by early-Universe data under the standard Λ CDM paradigm. Recent advances have identified the new Dark Energy Spectroscopic Instrument (DESI) survey as a promising observational tool for probing late-time cosmology. In this paper, we use the latest DESI DR2 baryon acoustic oscillation (BAO) measurements, combined with nonparametrically reconstructed Type Ia supernovae and cosmic chronometer data, to constrain the Hubble constant using three complementary machine learning methodologies: Gaussian process regression (GPR), artificial neural networks (ANNs), and long short-term memory (LSTM) networks. We perform individual and joint constraints on H 0 at the effective redshifts of DESI DR2 tracers, yielding H 0 = 61.8 plus or minus 5.3, km,s-1,Mpc-1 (GPR), 66.3 plus or minus 5.2, km,s-1,Mpc-1 (ANN), and 68.7 plus or minus 6.0, km,s-1,Mpc-1 (LSTM). Our results exhibit excellent agreement with the recent Planck measurements within 1σ, which highlight the potential of data-driven, model-independent methods to address the Hubble tension. Finally, we further test the sensitivity of our results to potential systematics in the DESI data, emphasizing the importance of understanding systematic uncertainties in BAO surveys.

Sources

Related papers