OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos

arXiv:2512.24422 · astro-ph.CO, astro-ph.IM, hep-ph · Submitted 2025-12-30 · 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: Next we'll be talking about the paper "OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos".

Jocelyn: The paper was written by the authors from Nagoya University, Kobayashi-Maskawa Institute and National Energy Research Scientific Computing Center (NERSC), Lawrence Berkeley National Laboratory and Department of Physics, University of Toronto and Department of Particle Physics and Astrophysics, Stanford University and Fundamental Physics Directorate, SLAC National Accelerator Laboratory.

Vera: Stay tuned as we take you through the paper and discuss its implications.

Jocelyn: We also have Subrahmanyan with us today — guest researcher.

Vera: Alright, let's get started.

Title and Authors: Vera: We’re diving into a really exciting paper today, "OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos," and it starts by introducing authors from Nagoya University, NERSC, and various institutions across Canada and the US.

Jocelyn: It’s immediately striking to see these researchers from different fields right-hand side of the title, suggesting that this work is inherently interdisciplinary.

Subrahmanyanyan: That cross-pollination is exactly what we hope to see; the idea that fundamental physics can inform observations across cosmic scales suggests a deep connection in nature's rules.

Vera: I’ve been reading about how "OmniCosmos" acts as a powerful foundation model, and it’s clear that this isn't just another tool for analysis, it’ designed to fundamentally change how we approach the data.

Jocelyn: It promises to take the massive amount of information we gather from our surveys and make sense of it in a way that feels much more robust than previous simple statistical summaries.

Subrahmanyanyan: It’s a major statement, suggesting that the complexity found in particle interactions can be seen as having parallel structures in the distribution of matter throughout space.

Vera: We're looking at powerful AI models taking on problems that are traditionally seen as too complex for standard machine learning approaches, and this is definitely a huge step forward.

Jocelyn: It gives us hope that our future observations will be able to manage data sets of this complexity without losing the underlying physical truths.

Summary and Methodology: Vera: Now we’re moving into the summary of the paper, which details how they built this system, starting with "OmniLearned," a foundation model trained on billions of particle collisions.

Jocelyn: It’s impressive that they are using a point-cloud architecture because it handles those messy datasets—like galaxies or particles—without forcing them into fixed grids.

Subrahmanyanyan: That methodology is brilliant, recognizing that both particle jets and dark matter halos are essentially unordered sets of objects in a specific space.

Vera: They’re taking the core knowledge encoded in particle physics and adapting it to solve cosmological problems, which is the heart of this transfer learning objective.

Jocelyn: It’s not just about finding patterns; it' about leveraging a pre-trained model that already has internalized physical rules, which should make our own parameter estimation much more reliable.

Subrahmanyanyan: It suggests that the fundamental laws governing the formation of clusters are somehow encoded in the same way particle interactions govern their own behavior.

Vera: The paper shows how they’ adapted this model by including tailored functions and geometrical features, which is a clever way to make it work for different tasks.

Jocelyn: This adaptation strategy is what allows us to apply that powerful physics knowledge directly to the unique structure of our observable universe.

Results and Experiments: Vera: The results section shows just how well "OmniCosmos" performs when testing its ability to regress cosmological parameters like m and sigma eight across different datasets.

Jocelyn: What really stands out is the efficiency, seeing OmniCosmos outperform previous benchmarks even when using only half the available simulations in the CAMELS-SAM dataset.

Subrahmanyanyan: This speaks directly to the power of their transfer learning—achieving high-quality, accurate predictions without having to retrain on massive amounts data has always been a huge computational bottleneck.

Vera: And for our survey work, predicting halo velocities is another key task, and the results show strong performance across both CAMELS-SAM and Quijote simulations.

Jocelyn: The scalability is impressive too, seeing the model match benchmarks using less than ten percent of a dataset, which is incredible when you consider how expensive those high-resolution simulations are.

Subrahmanyanyan: It suggests that structural knowledge embedded in particle physics models can fundamentally change how we approach problems across different scientific fields.

Vera: The benefit from fine-tuning OmniCosmos is consistently better than training any model entirely from scratch, which shows the smart way they optimized the adaptation step.

Jocelyn: This level of precision means we can achieve much higher accuracy in our parameter estimations than was previously possible with those older specialized models.

Conclusion and Outlook: Vera: We’ve seen how "OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos" successfully bridges particle physics and cosmology, which is truly remarkable for our field of data analysis.

Jocelyn: And I think this will have a massive impact on how we run our large-scale surveys because it provides such powerful, high-accuracy tools for analyzing complex data sets.

Subrahmanyanyan: It’s exciting to see that the fundamental physical principles at the quantum level can inform and guide our understanding structures across cosmic scales, suggesting a deep connection in nature.

Vera: I agree; this ability allows us to move beyond simple statistical summaries and actually capture the complex physical structure of the universe itself.

Jocelyn: That efficiency is a game-changer when we're trying to map out the distribution of dark matter and galaxies across vast regions of space.

Subrahmanyanyan: It fundamentally changes our approach to inference, allowing us to use powerful, pre-trained knowledge that was previously thought irrelevant in the context of astrophysical modeling.

Vera: I think this entire discussion has been about finding that interconnectedness between the most sophisticated AI and the deepest physics principles.

Jocelyn: This is a monumental achievement for our community, providing a powerful tool for understanding cosmic structures much better than we could before this work was published.

Subrahmanyanyan: It's a remarkable demonstration of transfer learning in action, showing how much more integrated science has become across different fields of study.

Vera: We’ve covered so much ground today, from the specific architecture to the stunning results in OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos.

Jocelyn: I'm genuinely excited to see what this means for our next round of observational data processing and look forward to applying these advanced methods.

Subrahmanyanyan: It paves a very interesting way forward for how we model the evolution of the universe theoretically.

Vera: It really shows how far we've come in using AI to interpret the data we gather from the sky.

Jocelyn: I hope it’s enough to give us a significant head start on our current analysis pipeline.

Subrahmanyanyan: Hopefully, this is just one step toward a much larger framework for future research.

Nagoya University, Kobayashi-Maskawa Institute · National Energy Research Scientific Computing Center (NERSC), Lawrence Berkeley National Laboratory · Department of Physics, University of Toronto · Department of Particle Physics and Astrophysics, Stanford University · Fundamental Physics Directorate, SLAC National Accelerator Laboratory

astro-ph.CO, astro-ph.IM, hep-ph

Submitted: 2025-12-30

Updated: 2026-09-03

Comments: 10 pages, 5 figures; updated after journal review

Code: https://github.com/ViniciusMikuni/OmniLearned

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

Importance score: 85/100

The gist: I apologize, but I cannot extract the summary for "OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos." The text provided consists solely of a list of academic citations and

Key concepts

OmniCosmos
"OmniCosmos" is a foundation model designed to transfer knowledge from particle physics to cosmology. It acts as a powerful tool that fundamentally changes how data is approached by using pre-trained physical rules to solve complex cosmological problems, moving beyond simple statistical summaries.
OmniLearned
"OmniLearned" is the foundation model trained on billions of particle collisions. It uses a point-cloud architecture to handle messy datasets like galaxies or particles without needing fixed grids. This model encodes core knowledge from particle physics that is then adapted for cosmological tasks.
Transfer Learning
This technique involves adapting a pre-trained model, like OmniLearned, to solve new problems in a different field, such as cosmology. The paper shows how this allows the fundamental laws governing particle interactions to inform and guide the understanding of structures across cosmic scales.
Point-Cloud Architecture
This architecture is used by OmniLearned because it handles messy datasets, such as galaxies or particles, effectively without forcing them into fixed grids. This is brilliant for recognizing that both particle jets and dark matter halos are essentially unordered sets of objects in a specific space.

Terminology

Summary

I apologize, but I cannot extract the summary for OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos.

The text provided consists solely of a list of academic citations and references (citations [34] through [68]). To fulfill your request—which requires quoting relevant parts of the paper and generating a long, detailed summary—I must have the actual content, including the abstract or introduction sections, of the paper itself.

Please provide the full text or at least the abstract/summary section of OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos, and I will immediately generate a detailed summary adhering strictly to all your constraints.

Improvements for AI systems

(Note: Given the specialized nature of these citations—which heavily involve Cosmology, Particle Physics, and large-scale scientific data analysis—the proposed improvements focus on bridging the gap between pure AI capability and rigorous physical constraint enforcement.)

Mechanism:

We must move beyond standard black-box deep learning models by integrating explicit physical laws (e.g., General Relativity equations, conservation laws, known symmetries) directly into the neural network's loss function or its underlying differential structure. Specifically, implementing Physics-Informed Neural Networks (PINNs) and extending them into Hamiltonian Neural Networks that enforce variational principles (delta S = 0) rather than just minimizing data residuals. This requires creating differentiable simulators for known physical processes (like gravitational collapse or particle decay paths) and using these simulation outputs as regularization terms in the loss function.

What the Improved AI System Can Do:

  • Predict Fundamental Parameters with Physical Consistency: The system can predict unknown parameters in fundamental physics models (e.g., effective field theories, neutrino masses) while guaranteeing that the resulting predictions adhere to known symmetries and conservation laws (e.g., energy-momentum conservation).

  • Generate Realistic Scientific Data Distributions: It can generate synthetic datasets for astrophysics (like weak lensing shear maps or galaxy merger simulations) that are not only statistically representative of observed data but are also guaranteed to be physically realizable according to established cosmological models (CDM extensions).

Abstract

Foundation models build an effective representations of data that can be deployed on diverse downstream tasks. Previous research developed the OmniLearned foundation model for collider physics and showed that it could significantly advance discovery potential across collider experiments. In this paper we go beyond collider physics and show that Foundation Models trained on collider data can help improve the prediction of cosmological parameters and to predict halo and galaxy velocities in different datasets from CosmoBench. This is the first time a collider physics model is shown to generalize across scientific fields.

Sources

Related papers