Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos

Vinicius Mikuni, Ibrahim Elsharkawy, Benjamin Nachman

TL;DR

The paper addresses transferring knowledge from collider-based foundation models to cosmology to improve parameter inference and halo velocity predictions from point-cloud data. It introduces OmniCosmos, a cosmology-adapted continuation of the OmniLearned model, with input-representation changes, geometry-aware pairwise features, and a carefully tuned fine-tuning regime (≈$2$ million parameters across $8$ transformer blocks) to handle up to ~5000 halos. Empirically, OmniCosmos delivers strong, data-efficient performance on CosmoBench tasks, surpassing scratch-trained baselines and matching or exceeding benchmarks on CAMELS-SAM and QUIJOTE with substantially fewer simulations. The work demonstrates cross-domain transferability of collider-pretrained foundation models to cosmology and points to future directions, including alternative adaptation strategies and generative capabilities for fast, high-fidelity cosmological surrogates.

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.

OmniCosmos: Transferring Particle Physics Knowledge Across the Cosmos

TL;DR

The paper addresses transferring knowledge from collider-based foundation models to cosmology to improve parameter inference and halo velocity predictions from point-cloud data. It introduces OmniCosmos, a cosmology-adapted continuation of the OmniLearned model, with input-representation changes, geometry-aware pairwise features, and a carefully tuned fine-tuning regime (≈ million parameters across transformer blocks) to handle up to ~5000 halos. Empirically, OmniCosmos delivers strong, data-efficient performance on CosmoBench tasks, surpassing scratch-trained baselines and matching or exceeding benchmarks on CAMELS-SAM and QUIJOTE with substantially fewer simulations. The work demonstrates cross-domain transferability of collider-pretrained foundation models to cosmology and points to future directions, including alternative adaptation strategies and generative capabilities for fast, high-fidelity cosmological surrogates.

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.
Paper Structure (5 sections, 2 equations, 5 figures, 3 tables)

This paper contains 5 sections, 2 equations, 5 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Knowledge Transfer
  3. CosmoBench Dataset
  4. Results
  5. Conclusion and Outlook

Figures (5)

  • Figure 1: Summary of the knowledge adaptation strategy between the OmniLearned model, trained on particle collisions, to cosmology tasks named OmniCosmos.
  • Figure 2: Results for the prediction of different cosmological parameters obtained using the CAMELS-SAM simulations.
  • Figure 3: Results for the halo velocity prediction using the CAMELS-SAM simulations.
  • Figure 4: Results for the prediction of different cosmological parameters obtained using the QUIJOTE simulations.
  • Figure 5: Results for the halo velocity prediction using the QUIJOTE simulations.