Mapping the Universe with Galaxy Family Trees

by

Author: Denis Avetisyan

New research shows how machine learning can connect the structure of cosmic history to the properties of dark matter and galaxy evolution.

Within a cosmological simulation, the interwoven history of a Milky Way-mass subhalo—traced from its earliest moments to the present—demonstrates how even vast structures are built from the merging of smaller components, their masses ranging from 7x10<sup>11</sup> to 2.5×10<sup>12</sup> solar masses, a humbling reminder that even the grandest constructions are subject to dissolution and recombination.” style=”background:#FFFFFF” /><figcaption>Within a cosmological simulation, the interwoven history of a Milky Way-mass subhalo—traced from its earliest moments to the present—demonstrates how even vast structures are built from the merging of smaller components, their masses ranging from 7×10<sup>11</sup> to 2.5×10<sup>12</sup> solar masses, a humbling reminder that even the grandest constructions are subject to dissolution and recombination.</figcaption></figure> <p><b>Graph Neural Networks trained on merger trees from cosmological simulations successfully infer both warm dark matter particle mass and astrophysical feedback parameters.</b></p> <p>The standard cosmological model relies on parameters difficult to constrain, yet galaxy formation is intimately linked to underlying dark matter physics. This research, <i>‘Linking Warm Dark Matter to Merger Tree Histories via Deep Learning Networks’</i>, explores whether the hierarchical growth of structure, represented by galaxy merger trees, can be used to infer cosmological and astrophysical parameters. We demonstrate that Graph Neural Networks trained on these merger tree histories can effectively predict warm dark matter particle masses and supernovae feedback parameters with surprising accuracy. Could this approach offer a novel pathway to simultaneously constrain both cosmological models and the complex processes governing galaxy evolution?</p> <hr> <h2>Echoes of Creation: The Foundations of Cosmic Structure</h2> <p>Understanding galaxy formation requires untangling the interplay of dark matter, gas, and stellar processes. Cosmological simulations reveal galaxies emerge within a hierarchical network of merging dark matter halos, seeded by early density fluctuations amplified by gravity. Feedback from supernovae and active galactic nuclei profoundly shapes galaxy evolution; supernovae regulate star formation, while active galactic nuclei suppress it. The nature of dark matter – cold or warm – significantly alters predicted structures; cold dark matter favors bottom-up formation, while warm dark matter suppresses small-scale structures. Each elegant model, however, remains a fragile construction, susceptible to collapse under new evidence—a provisional map drawn on shifting sands.</p> <figure> <img decoding=
Analysis of feature importance in an XGBoost model reveals that the number of nodes is more than twice as influential as any other feature, indicating a strong sensitivity to wavelength division multiplexing, while gas-to-dark matter mass ratio and stellar mass also contribute significantly to the model’s predictions.

Simulating the Cosmos: A Computational Mirror

Cosmological simulations are indispensable tools for understanding the universe and large-scale structure. Projects like the DREAMS Project generate vast datasets modeling the gravitational interactions of matter over cosmic timescales. These simulations rely on accurately representing both dark matter gravity and gas hydrodynamics, utilizing N-body techniques and sophisticated algorithms to solve for radiative cooling, star formation, and feedback. A key focus is the hierarchical merging history of dark matter halos, providing a natural explanation for the observed distribution of galaxies and clusters, and connecting galaxy properties to the early universe.

Graphing the Darkness: Unlocking Simulation Insights

Graph Neural Networks (GNNs) offer a robust methodology for analyzing complex relationships within cosmological simulation data, particularly leveraging merger trees that detail the hierarchical growth of structures. This allows for a nuanced understanding of galaxy formation and evolution. GNNs accurately infer crucial galaxy properties – stellar mass, gas mass, and star formation rate – from simulation data, incorporating dark matter mass and simulation epoch. Quantitative analysis reveals a strong predictive power, with an R2 value of 0.957 when predicting Warm Dark Matter particle mass, suggesting GNNs are valuable for constraining cosmological parameters.

A tree-to-graph conversion process, coupled with a graph neural network architecture, encodes node features and edges to facilitate message passing through multiple layers and subsequent aggregation, ultimately producing a predictive output.
A tree-to-graph conversion process, coupled with a graph neural network architecture, encodes node features and edges to facilitate message passing through multiple layers and subsequent aggregation, ultimately producing a predictive output.

Beyond Prediction: Glimpses of Cosmic Truth

Cosmological simulations, combined with GNNs, provide a novel pathway for analyzing galaxy formation and evolution, extracting insights from complex relationships. Recent studies demonstrate the effectiveness of this approach in predicting key galactic parameters. GNN models ASN1A and ASN2A achieve high accuracy in predicting supernova wind energy and velocity (R2 of 0.987 and 0.880, respectively), indicating a strong correlation between simulated data and predictions.

Comparison of inferred versus true parameter values demonstrates that the ASN1A and ASN2A graph neural network models achieve high accuracy (R² of 0.987 and 0.880, respectively), while the AAGNA model exhibits poor performance (R² of 0.102).
Comparison of inferred versus true parameter values demonstrates that the ASN1A and ASN2A graph neural network models achieve high accuracy (R² of 0.987 and 0.880, respectively), while the AAGNA model exhibits poor performance (R² of 0.102).

Future research will expand these techniques to encompass larger, more intricate simulations, refining models and pushing the boundaries of knowledge. Each iteration brings us closer to the invisible truth, though, like the event horizon itself, it forever recedes from our grasp.

The research detailed within this article underscores the inherent limitations of predictive modeling in complex systems, echoing Grigori Perelman’s sentiment: “It is better to be skeptical than to believe.” Just as a black hole’s event horizon obscures information, the vastness of cosmological simulations and the interplay of warm dark matter and astrophysical feedback introduce uncertainties. The application of Graph Neural Networks, while offering a powerful tool to infer parameters from merger tree histories, doesn’t circumvent the fundamental challenge of incomplete information. Gravitational lensing around a massive object allows indirect measurement of black hole mass and spin; similarly, this study utilizes network analysis to deduce cosmological properties, acknowledging that any attempt to predict object evolution requires numerical methods and Einstein equation stability analysis. The inferred parameters, therefore, represent probabilistic estimations, not absolute truths.

What’s Next?

The capacity to infer cosmological and astrophysical parameters from the tangled histories of galactic mergers – as demonstrated by this work – feels less like a triumph of understanding and more like a sophisticated form of pattern recognition. It is a testament to the power of these graph neural networks, certainly, but also a reminder that correlation does not equate to causation. When light bends around a massive object, it’s a reminder of limitations. The true nature of dark matter, warm or otherwise, remains elusive, hiding behind layers of inference.

Future work will undoubtedly refine these networks, increasing their complexity and incorporating even more detailed simulations. Yet, the fundamental question lingers: are these models truly revealing the universe, or simply reflecting the biases embedded within the simulations themselves? They are like maps that fail to reflect the ocean. The incorporation of more realistic astrophysical feedback mechanisms is crucial, but even then, the inherent uncertainties in these processes will continue to cast a long shadow.

Perhaps the most fruitful avenue for future research lies in confronting these learned inferences with independent observational constraints. The ultimate test will not be the ability to reproduce simulated histories, but to predict the behavior of galaxies yet unseen. For every answer obtained, a multitude of new questions will inevitably emerge, drawing the field ever deeper into the darkness – a darkness that, after all, may be the truest reflection of its own ambition.

Original article: https://arxiv.org/pdf/2511.05367.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-11-11 06:59