🔭 astrophysics

The EPS Research Astro-RAG Platform: A Unified Open-Science Infrastructure for Cross-Epoch Astrophysical Kinematic Analysis, LLM-Assisted Research Workflows, and Educational Outreach

The EPS Research Astro-RAG Platform (v1.0) is an open-science infrastructure that unifies four cross-epoch astrophysical corpora covering 772 objects from the local universe to high-redshift galaxies, revealing a robust sign reversal in angular velocity from positive values at z=0 to negative values at z~5, while providing executable Jupyter notebooks, educational tracks, and a roadmap for LLM-assisted research workflows.

Original authors: David C. Flynn

Published 2026-06-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: David C. Flynn

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the universe as a massive, multi-generational family album. For decades, astronomers have been taking photos of this family at different ages: some pictures show the family as babies (distant, ancient galaxies), others show them as teenagers (our local neighborhood), and some show them as adults (nearby stars).

The problem is that everyone took these photos with different cameras, in different languages, and stored them in different albums. If you wanted to compare how the family moved from their baby phase to their adult phase, you'd have to spend years just translating the photos and organizing the albums.

The EPS Research Astro-RAG Platform is like a new, super-organized digital library that fixes this mess. It takes all those scattered photos and organizes them into one single, easy-to-read book.

Here is what this "library" actually does, broken down simply:

1. The Four Main Collections (The "Albums")

The platform gathers 772 different cosmic objects into four specific groups, all speaking the same language:

The Local Neighborhood: 438 nearby galaxies (like our own Milky Way) where we can see exactly how they spin.
The "Little" Galaxies: 129 tiny, irregular galaxies that are mostly made of dark matter.
The Star Clusters: 174 dense groups of stars (globular clusters) right here in our galaxy.
The Ancient Ancestors: 31 very old, distant galaxies from when the universe was just a toddler (about 12 billion years ago).

2. The "Universal Translator" (The Data Format)

Usually, scientific data is like a secret code that only experts can read. This platform translates everything into a format that both humans and computers (specifically, advanced AI) can understand instantly.

Think of it as taking a handwritten letter in a foreign language and turning it into a clear, digital text file that anyone can search, copy, and paste.
They even created 120 "practice guides" (computer programs called Jupyter notebooks) that show you exactly how to use this data, from simple charts to complex math.

3. The Big Discovery: The "Spin Reversal"

The most exciting part of this paper is a specific discovery they made by comparing the "baby" galaxies to the "adult" galaxies.

They measured something called omega ( $\omega$ ), which is basically a way to describe how fast the edges of a galaxy are spinning compared to its center.

The "Adult" Galaxies (Today): When they looked at nearby galaxies, the spin value was positive (like a clock moving forward).
The "Baby" Galaxies (Ancient Times): When they looked at the ancient galaxies, the spin value was negative (like a clock moving backward).

The Analogy: Imagine a figure skater. In the ancient past (high redshift), they were spinning in one direction. Over billions of years, as the universe evolved, they somehow flipped and started spinning in the opposite direction. The platform proves this "flip" happened across 9 billion years of cosmic history.

4. Making Science Accessible (The "School Track")

This isn't just for professional scientists. The platform includes a special "High-School Track."

They took the same real, complex data used by experts and built 20 simple lessons for students aged 12–18.
No prior knowledge needed: A 12-year-old can use these tools to see real galaxies and understand how they move, without needing a degree in physics first.

5. The Future: AI Helpers

The authors are building a bridge to the future. They are designing this data specifically so that Artificial Intelligence (AI) can read it easily.

In the next phase, they plan to train AI models to act as research assistants. These AIs will be able to look at a picture of a galaxy and instantly tell you how fast it's spinning, or answer questions about the data without a human having to do the heavy lifting.

Summary

In short, this paper presents a universal toolkit for studying how galaxies move and change over time. It organizes messy data, proves that galaxies used to spin differently than they do today, and hands the keys to this discovery to everyone—from professional researchers to high school students—so they can explore the universe together.

Technical Summary: The EPS Research Astro-RAG Platform

Problem Statement
The paper identifies a critical bottleneck in modern astrophysics: while large-scale surveys (e.g., SPARC, THINGS, LITTLE THINGS, WALLABY, ALPINE) have generated rich kinematic datasets spanning vast ranges in galaxy mass and cosmic epoch, these data remain trapped in heterogeneous formats across dozens of source papers. This fragmentation hinders systematic cross-survey and cross-epoch analysis. Furthermore, the emergence of Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) pipelines introduces a new requirement: datasets must be not only machine-readable but "self-describing" (structured with embedded units and metadata) to allow LLMs to ingest and reason over data without external preprocessing. Standard FITS tables and survey-specific ASCII formats fail to meet this criterion.

Methodology and Platform Architecture
To address these challenges, the authors present the EPS Research Astro-RAG Platform (v1.0), a unified open-science infrastructure designed to bridge raw data, scientific analysis, and AI-assisted workflows. The platform is organized into five functional "silos":

z = 0 Data: Local universe corpora (HI galaxies, dwarfs, globular clusters).
z ∼6 Data: High-redshift kinematic data.
Example Library: 120 verified Jupyter notebooks.
Papers & Preprints: Publication metadata.
LLMs & Tools: Planned fine-tuned model adapters.

The core methodology involves distributing four distinct corpora under a common schema available in four formats: hierarchical JSON (self-describing), flat CSV (for pandas/R), JSONL (optimized for LLM RAG streaming), and per-object ZIP files. Quality tiers are explicitly defined (Tier 1 for full kinematic data, Tier 2 for integrated parameters or classifications only).

Key Contributions
The platform releases four specific corpora containing a total of 772 objects:

Unified HI Rotation Curve Corpus v7.0: Contains 438 spatially resolved rotation curves from SPARC, THINGS, LITTLE THINGS, and WALLABY DR2. It preserves the full fidelity of the SPARC database, including all photometric decomposition components ( $V_{gas}$ , $V_{disk}$ , $V_{bul}$ ) and surface brightness profiles, augmented with data from other surveys.
Dwarf/Irregular HI Corpus v1.0: Contains 129 low-mass, dark-matter-dominated galaxies from LVHIS, VLA-ANGST, LITTLE THINGS, and WALLABY.
Milky Way Globular Cluster Corpus v1.3.1: Contains 174 clusters harmonized from the Harris catalog, Gaia EDR3 proper motions, HST structural parameters, and N-body dynamical models.
High-z Kinematic Corpus Z1: Contains 31 star-forming galaxies at $z = 4.26–5.68$ from the ALPINE survey, including 8 confirmed 3DBarolo rotators.

The platform also includes a High-School Exploration Track (20 notebooks) designed to make these real, published datasets accessible to students aged 12–18 without prior domain knowledge, and a QuickStart pathway that allows users to reproduce core results in under 10 minutes.

Scientific Results
The unifying scientific contribution across these corpora is the application of the omega kinematic correction (Flynn & Cannaliato 2025). This method introduces a boundary-point angular velocity parameter ( $\omega$ ) derived from the innermost and outermost resolved rings of rotation curves.

Local Universe ( $z=0$ ): Applied to 84 SPARC quality-1 galaxies, the correction yields a mean $\omega = +7.06 \pm 3.26$ rad Gyr $^{-1}$ . The dwarf/irregular corpus yields a median $\omega = +9.94$ rad Gyr $^{-1}$ .
High Redshift ( $z \sim 5$ ): Applied to 8 confirmed rotators in the Z1 corpus (using ALMA [CII] 3DBarolo models), the parameter yields a median $\omega = -13.05$ rad Gyr $^{-1}$ .
Sign Reversal: The platform demonstrates a robust sign reversal of $\sim 20$ rad Gyr $^{-1}$ across $\sim 9$ Gyr of cosmic evolution, contrasting positive values at $z=0$ with negative values at $z \sim 5$ .

Significance and Claims
The paper claims the platform provides the first unified kinematic dataset connecting HI 21cm rotation curves, Milky Way globular cluster dynamics, and high-redshift ALMA [CII] morpho-kinematics under a common schema.

Reproducibility: The authors emphasize that all 120 Jupyter notebooks are verified to execute without errors (120/120 pass rate) against the actual Zenodo corpus files, ensuring valid data access patterns.
Educational Access: By providing a "High-School Exploration Track" that loads directly from the same Zenodo files used by researchers, the platform claims to democratize access to real astrophysical data, removing the need for proprietary software or telescope archive accounts.
Future Roadmap: The sign reversal in $\omega$ is presented as a motivation for future RAMSES cosmological simulations (planned for 2026–2027) to trace this structural transformation. Additionally, the platform outlines a "Stage 2" plan to deploy a four-tier fine-tuned LLM team (ranging from ~7B to ~72B parameters) specifically trained on these corpora for RAG pipelines, including a vision-capable model to extract kinematic values directly from plots.

The authors state that the platform is released under CC BY 4.0 with permanent DOIs, ensuring open access and long-term archival stability.