Testing CDM with ANN-Reconstructed Expansion History from Cosmic Chronometers

arXiv:2604.22372 · astro-ph.CO · Submitted 2026-04-24 · 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: "Testing CDM with ANN-Reconstructed Expansion History from Cosmic Chronometers".

Jocelyn: In modern cosmology, this work proposes a novel artificial neural network (ANN)-based framework for the non-parametric reconstruction of late-time cosmic expansion using Cosmic Chronometer data.

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

Paper summary: Vera: So, looking at the "Testing CDM with ANN-Reconstructed Expansion History from Cosmic Chronometers" paper by Hashimoto et al., what stands out is their commitment to building a reproducible methodology for this kind of reconstruction work. They’ve taken a complex task and broken it down into distinct stages to ensure that implementation choices are thoroughly vetted before they ever touch real data.

Jocelyn: I think the authors really drive home the point that their framework isn't just about getting one good result, but establishing a system—a pipeline—where you can see exactly how each step contributes to the final output, which is pretty important for any scientific method.

Subrahmanyan: From a theoretical perspective, this work suggests that ANN architectures are viable tools for mapping complex relationships in cosmological data like expansion history, even when we are trying to test standard models like ΛCDM one. The ability of the framework to align with ΛCDM predictions within observational uncertainties is a strong indicator of its potential utility.

Vera: That alignment is certainly a major finding, showing that this reconstruction method can yield results consistent with what we already expect from the standard cosmological model, which reinforces confidence in the technique itself.

Jocelyn: And they make it clear that the external H0 priors they use are an explicit part of their Stage three setup, meaning you have to account for those initial conditions when interpreting the final Hubble parameter predictions, as shown by their comparisons across Planck two thousand eighteen TRGB, and SH0ES R21.

Subrahmanyan: So the main implication is that we can use this ANN-based framework to test ΛCDM consistency in a way that allows us to explicitly quantify the effect of different cosmological anchors on the reconstructed expansion history

six–twelve: .

Vera: It gives us a new avenue for testing the standard model, moving toward more flexible reconstruction strategies instead of being strictly limited by low-dimensional parametric assumptions.

Jocelyn: Ultimately, this paper provides a structured approach to using AI tools in cosmological inference that prioritizes validation through mock data before applying it to the actual Cosmic Chronometer observations.

Conclusion: Vera: So, we’ve been going through the technical weeds of this paper on reconstructing cosmic expansion history using an AI framework called Cosmic Chronometer data, and now we need to step back and look at what it actually means for us as an observational community.

Jocelyn: Exactly, Vera. That title itself is really telling; "Testing CDM with ANN-Reconstructed Expansion History from Cosmic Chronometers" tells me this isn't just some abstract mathematical exercise in a lab; it’s directly aimed at checking our current cosmological model against real sky data we already have.

Subrahmanyan: From a theoretical standpoint, the real interest here is how this AI handles the non-parametric reconstruction of expansion history. It suggests that complex, messy observational data can be mapped onto a smooth expansion curve without being strictly limited by simple parametric assumptions like those in ΛCDM.

Vera: I’m seeing how they use that staged methodology—separating mock validation from real data application—as the most important part for me as an astronomer looking at the sky; it shows they're not just throwing a black box at the data without testing if it actually makes sense first.

Jocelyn: And I agree, Vera, that rigorous process of testing implementation choices before touching the actual CC data is what gives this work its credibility in my field; it’s about building trust in the output from the start.

Subrahmanyan: That level of structural validation is crucial because if we use these kinds of tools to probe cosmology, we need to know precisely how much confidence we can place in the resulting expansion history. The paper’s results showing alignment with existing models under certain priors give us a solid anchor point.

Vera: It really does, and those specific results they got for the Hubble parameter predictions across different external anchors—Planck two thousand eighteen TRGB, SH0ES R21—that’s what makes this paper compelling for me; it shows consistency even when you change your starting point.

Jocelyn: And that explicit dependence on those external priors in Stage three is a key observation for me as someone who deals with anchor measurements like the Hubble constant; it confirms that the prior itself plays a significant role in what the AI spits out.

Subrahmanyan: So, essentially, this work points toward using sophisticated machine learning not just to fit parameters into a pre-defined model, but to explore the space of possible expansion histories constrained by real observational data.

Vera: That’s the high-level idea: using the power of AI to reconstruct cosmology in a way that respects both the complexity of real data and our existing theoretical frameworks.

Jocelyn: It opens up a new avenue for testing how well our current cosmological assumptions hold up when we use these advanced reconstruction methods instead of just relying on traditional fitting techniques.

Subrahmanyan: Moving forward, the implication is that we might see AI-driven tools used more frequently to explore the boundaries of what’s possible in cosmic expansion measurements.

Vera: It’s definitely an exciting direction for observational cosmology, and I think we need to keep watching how these staged reconstruction methods evolve.

Faculty of Symbiotic Systems Science, Fukushima University

astro-ph.CO

Submitted: 2026-04-24

Updated: 2026-10-01

Comments: 33 pages, 8 figures, version accepted for publication in Journal of Cosmology and Astroparticle Physics

Journal ref: JCAP 09, 146 (2026)

DOI: 10.1088/1475-7516/2026/09/146

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: In modern cosmology, this work proposes a novel artificial neural network (ANN)-based framework for the non-parametric reconstruction of late-time cosmic expansion using Cosmic Chronometer data.

Key concepts

Cosmic Chronometer (CC) Data
These are observational data points that provide constraints on the expansion history of the universe at late times. The study uses these specific data to train and test an AI model designed to predict how the universe has expanded over time.
Staged Methodology
The research follows a three-stage process: first, creating mock data for validation; second, screening different network architectures and activation functions based on performance metrics; and third, training the final model on real data. This structured approach ensures reproducible results by fixing implementation choices before applying them to actual observations.
Weighted L1 Objective
This is the mathematical function used to train the neural network. It penalizes large errors between the model's prediction and the true target value, weighted by a factor related to observational uncertainty. This loss function guides the ANN in learning an expansion history that fits both the data and accounts for measurement noise.
H0 Prior Dependence
The final reconstruction results are sensitive to which external measurement of the Hubble constant (H0) is used as an input anchor. The study shows this dependence explicitly, meaning the choice of H0 prior must be considered a necessary part of the final analysis setup.

Terminology

Summary

In modern cosmology, this work proposes a novel artificial neural network (ANN)-based framework for the non-parametric reconstruction of late-time cosmic expansion using Cosmic Chronometer data. The research matters because it establishes a reproducible, staged methodology—separating mock-based validation from real-data application—to test the robustness of ANN approaches in probing the expansion history, aligning reconstructed results with existing cosmological models like ΛCDM.

The gist

The framework is formulated as a reproducible reconstruction pipeline in which the implementation choices themselves are part of the methodology, separating mock-based validation and architecture selection from the final fit to observed CC data.

Methodology and Staged Workflow

The framework operates through a three-stage screening pipeline designed to fix implementation choices before applying them to real data. Stage 1 involves Mock data generation, where a suite of CC-like samples is created from a fiducial flat-ΛCDM background, used only for controlled validation and design selection. The mock redshifts are drawn from a Gamma distribution whose parameters are fixed by moment matching, and the mock uncertainty model is calibrated by fitting linear trends to the reported CC uncertainties.

Stage 2 focuses on Candidate ANN training/ Activation screening and Architecture selection. This stage involves:

  1. Comparing candidate activation functions (e.g., ELU, tanh, SiLU) under a common training objective to rank them based on mock-truth recovery and smoothness diagnostics.

  2. Scanning network width/depth for the preferred activation function to select the best hyperparameter set. The ranking statistic is defined as Score = Sselect + λsmoothRsmooth, where Sselect measures truth-recovery error in units of the mock observational uncertainty and Rsmooth penalizes unnecessarily oscillatory reconstructions.

Network Architecture and Loss Functions

The baseline reconstructor is a feed-forward multilayer perceptron mapping redshift directly to the Hubble parameter, with input redshift normalized as xi = zi − zloc / zscale. The network architecture search compares models with varying widths from 8 to 256 neurons. The loss function implemented throughout the design study and final reconstruction is the weighted L1 objective, defined as:

LwL1 = (1/N) Σ ŷi − yisi + εs, where ŷi is the normalized ANN prediction, yi is the normalized target, si is the normalized observational uncertainty, and εs is a small numerical floor.

Final Reconstruction and Evaluation

Stage 3 involves Observed CC data training, where the selected architecture (the exported single-hidden-layer [8] model) is trained on the real CC compilation augmented by an external H0 prior at z = 0. The final prediction is constructed from an explicit ensemble of 100 independently initialized training members. This ensemble mean prediction, denoted as H¯(k)(zi), is used for comparison across three different external H0 priors: Planck 2018, TRGB, and SH0ES R21.

Results and Conclusions

The results show that the exported Stage-2 model yields a smooth reconstruction for all three archived H0 priors. The ANN predictions at z = 0 are HANN0 = 67.412 for the Planck 2018 anchor, 67.240 for the TRGB anchor, and 73.239 for the SH0ES R21 anchor, demonstrating that the prior dependence is explicit and should be regarded as part of the Stage-3 setup. The empirical diagnostics across all three runs remain stable, with RMSE values ranging from 11.45 to 14.18 km s−1 Mpc−1 and empirical 68% coverages between 0.853 and 0.941, confirming the reliability of this structured pipeline for ANN-based cosmological reconstruction. The main contribution is establishing a reproducible staged baseline grounded in a documented mock-based selection procedure to improve reproducibility over loosely constrained tuning methods.

Current Limitations and Future Extensions

Limitations include treating only a CC-only branch supplemented by one external H0 anchor, which limits testing the ANN design logic with covariance-rich multi-probe likelihoods. Furthermore, the architecture is optimized for a smooth fiducial flat-ΛCDM background rather than reconstructions with sharp localized features. Future work includes propagating this staged design logic to multi-probe forward models involving Pantheon+ and BAO observables, comparing the current sigma-head plus ensemble prescription with fully covariance-aware objectives, and extending the benchmark strategy to derived quantities such as E(z), Om(z), and numerical derivatives of H(z).

Improvements for AI systems

As a fastidious and diligent researcher, I have analyzed the proposed ANN-based framework for non-parametric reconstruction of cosmic expansion history using Cosmic Chronometer (CC) data. The core innovation lies in the three-stage pipeline: mock generation, architecture/activation screening (Stage 2), and real-data reconstruction augmented by external priors (Stage 3).

Here are specific improvements to AI systems derived from this paper, along with what those improved systems can do:

)

)

  1. [Improved System: The Staged, Prior-Aware Reconstruction Engine]

Based on the explicit separation of mock-based design and real-data fitting (Section C), the system can be engineered to perform robust model selection before committing computational resources to complex data. By fixing the activation function and architecture based on a truth-aware score (Equation 2.32) against mock data, the system avoids wasting training time on sub-optimal network structures or non-smooth functions when applied to real data.

)

The improved system can:

A. Perform highly efficient, pre-validated reconstruction of cosmological expansion history from new observational datasets (like future CC or novel probes).

B. Automatically select the optimal neural network architecture (e.g., identifying that a shallow [8] network is superior to deeper ones for a given data regime) based on performance metrics derived from simulated truth.

C. Produce reconstructions where the uncertainty band explicitly combines two distinct sources: the internal aleatoric uncertainty predicted by an integrated sigma-head and the ensemble dispersion of multiple independently initialized network members (Equation 2.37).

  1. [Improved System: Robust Prior-Conditioned Inference Module]

The framework demonstrates how to integrate external, low-redshift constraints (like Planck or TRGB) into a non-parametric ANN reconstruction by treating them as Gaussian priors at the anchor point z=0 (Section C and Table V). The system can be specifically tuned to handle these prior-augmented inputs.

The improved system can:

A. Provide prior-conditioned reconstructions of the late-time expansion history, allowing researchers to test how different low-redshift constraints (e.g., Planck vs. SH0ES) influence the reconstructed dark energy evolution across the entire redshift range.

B. Quantify the sensitivity of the reconstruction to specific anchor priors by analyzing shifts in H(z) at z=0 and changes in derived cosmological parameters like Om(z).

  1. [Improved System: Automated Feature Extraction for Null Tests]

The paper emphasizes that derivatives, such as the expansion rate derivative H'(z), are crucial for null tests and model comparison (Section VI). The framework is designed to be extended to compute these directly from the reconstructed function.

The improved system can:

A. Calculate high-fidelity numerical derivatives of the reconstructed Hubble parameter, enabling rigorous testing of cosmological models against observational data (e.g., distinguishing between a flat-LambdaCDM model and a slightly tilted dark energy evolution).

B. Generate diagnostics like Om(z) [Equation 3.9] with propagated uncertainty bands, which are essential for comparing the shape of the expansion history against theoretical predictions derived from modified gravity theories or alternative dark energy models.

  1. [Improved System: Adaptive Loss Function and Uncertainty Modeling]

The system utilizes a weighted L1 loss (Equation 2.29) and explicitly models uncertainty through a sigma head (Equation 2.28), which is then combined with ensemble dispersion in Stage 3 (Equation 3.7).

The improved system can:

A. Be adapted to use covariance-aware loss functions, moving beyond simple L1 to incorporate complex, multi-probe error correlations that are currently only discussed as future extensions (Section VI).

B. Generate sophisticated uncertainty bands that accurately reflect both the model's inherent noise (via the sigma head) and the variance introduced by ensemble sampling across different initializations.

Sources

Related papers