🔭 astrophysics

A Unified HI Rotation Curve Corpus for Computational Astrophysics: 438 Galaxies from SPARC, THINGS, LITTLE THINGS, and WALLABY DR2

This paper presents a unified, publicly available corpus of 8,963 spatially resolved HI rotation curve measurements across 438 galaxies from four major surveys, structured in JSON and CSV formats with a two-tier quality system to facilitate both traditional numerical analysis and Large Language Model retrieval-augmented generation pipelines.

Original authors: David C. Flynn

Published 2026-04-16

📖 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 giant library. For decades, astronomers have been collecting books about how galaxies spin. These "books" are called rotation curves, and they tell us how fast stars and gas are moving at different distances from the center of a galaxy.

Why does this matter? Because when we look at these spinning galaxies, they spin way faster than the visible stuff (stars and gas) should allow. This is the biggest clue we have that Dark Matter exists—a mysterious, invisible substance holding galaxies together.

However, there's a problem. The library is messy.

One group of researchers (SPARC) put their books on a wooden shelf.
Another group (THINGS) put theirs in a metal filing cabinet.
A third group (LITTLE THINGS) used a digital tablet.
A fourth group (WALLABY) just threw them in a cardboard box.

Each group used different units (some measured in miles, some in kilometers), different file formats, and different labels. If a scientist wanted to compare all these books to solve the Dark Matter mystery, they had to spend years just translating the languages and re-shelving the books. It was a nightmare for computers trying to read them automatically.

The Solution: The "Universal Galaxy Translator"

This paper introduces a new project by David Flynn that acts like a super-librarian. He has taken all those scattered books from the four major surveys and organized them into a single, perfectly structured digital collection.

Here is what he did, using some simple analogies:

1. The Great Unification (The "Master Menu")
Think of the four surveys as four different restaurants, each serving a delicious but slightly different version of the same dish (galaxy rotation data).

SPARC is the fancy restaurant with a full menu (including details on the "baryonic" ingredients like gas and stars).
THINGS and LITTLE THINGS are high-end spots with very precise measurements.
WALLABY is a massive fast-food chain that serves thousands of galaxies quickly, but with slightly less detail on the ingredients.

Flynn has created a Master Menu (a single computer file) that lists every dish from every restaurant in the exact same format. Now, a computer can order a "galaxy meal" from any of these places without getting confused by the language.

2. The Quality Tags (The "Star Rating")
Not all data is created equal. Flynn added a two-tier quality system, like a hotel rating:

Tier 1 (The 5-Star Hotels): These are the SPARC, THINGS, and LITTLE THINGS galaxies. They were hand-checked by human experts. Every single data point has a "safety check" (uncertainty numbers) attached to it.
Tier 2 (The Budget Chains): These are the WALLABY galaxies. They were processed automatically by a robot pipeline. They are still useful and reliable for big trends, but they don't have the same level of detailed "safety checks" for every single point.

3. The "Smart" Format (The "AI-Ready" File)
The most exciting part of this paper is that Flynn didn't just organize the data for humans; he organized it for Artificial Intelligence (AI).

Imagine you ask a robot assistant: "Show me the spinning speed of Galaxy DDO 161 and calculate its invisible mass."
In the past, the robot would get lost trying to find the right file.
With this new corpus, the robot can instantly find the "DDO 161" file, read the data, and even write the code to draw the graph for you. The authors tested this with four different AI models, and they all succeeded in understanding the data without needing a manual.

Why is this a big deal?

Speed: Scientists can now analyze thousands of galaxies in minutes instead of months.
Accuracy: By putting everything in one place with consistent units (kilometers per second and kiloparsecs), we can compare apples to apples, not apples to oranges.
The Future: This is a "training ground" for the next generation of AI astronomers. It teaches computers how to read galaxy data so they can help us discover new physics about Dark Matter and how the universe works.

In short: David Flynn took a chaotic pile of galaxy data, cleaned it up, labeled it clearly, and packed it into a box that both humans and AI robots can open and use immediately. It's a massive step forward in our quest to understand the invisible universe.

1. Problem Statement

Despite the critical role of galaxy rotation curves in understanding the mass discrepancy problem (dark matter vs. modified gravity), existing data is highly fragmented. Four major HI surveys—SPARC, THINGS, LITTLE THINGS, and WALLABY DR2—provide the bulk of high-quality rotation curve data for nearby galaxies. However, these datasets suffer from:

Heterogeneity: They are distributed across incompatible platforms with different file formats.
Inconsistent Conventions: Discrepancies in units (arcseconds vs. kiloparsecs), velocity definitions, and column naming.
Lack of Standardization: No unified, machine-readable schema exists to combine these surveys, creating a significant barrier for reproducible computational analysis, automated fitting pipelines, and emerging Large Language Model (LLM) Retrieval-Augmented Generation (RAG) workflows.
Manual Overhead: Researchers must individually download, reconcile, and verify data, which is inefficient and prone to error.

2. Methodology

The author developed the Unified Galaxy HI Rotation Curve Corpus (v7.0) through a rigorous ingestion and verification process:

Data Aggregation: The corpus aggregates 8,963 spatially resolved HI rotation curve measurements across 423 galaxies (plus 15 metadata-only THINGS entries), totaling 438 catalog entries.
- SPARC: 175 galaxies (Tier 1).
- THINGS: 34 entries (19 with full kinematic data, 15 metadata-only).
- LITTLE THINGS: 26 dwarf irregulars (Tier 1).
- WALLABY DR2: 203 galaxies from automated pipelines (Tier 2).
Verification & Correction:
- Kinematic parameters (inclination, distance, position angle) were cross-verified against scanned primary tables (Lelli et al. 2016; de Blok et al. 2008; Oh et al. 2015).
- A systematic check identified and corrected a unit conversion bug (missing $\times 1000$ factor) in an earlier ingestion iteration.
- WALLABY data was ingested from CADC 3DBarolo model files and cross-matched with the DR2 kinematic catalogue.
Standardization:
- Units: All radii converted to kiloparsecs (kpc); all velocities to km/s.
- Baryonic Decomposition: For SPARC, baryonic velocities ( $V_{gas}, V_{disk}, V_{bul}$ ) are stored at a mass-to-light ratio of $\Upsilon=1$ . A sign-preserving quadrature convention is used to handle negative gas velocities in inner disks.
- Quality Tiers: A two-tier system was implemented:
  - Tier 1: Hand-curated, per-point uncertainties, verified parameters (SPARC, THINGS, LITTLE THINGS).
  - Tier 2: Automated pipeline products without per-ring uncertainties (WALLABY DR2).

3. Key Contributions

The paper introduces a structured, multi-format data release designed for both traditional numerical analysis and modern AI workflows:

Unified Schema: A single JSON master file (rotation_curve_corpus_v7.json) containing nested per-ring kinematic data, survey metadata, and quality annotations.
Dual-Format Distribution:
1. Flat CSV: 438 rows for rapid catalog-level filtering and sample selection.
2. Per-Galaxy ZIP: 438 individual JSON files optimized for LLM RAG ingestion, where each galaxy acts as a self-contained retrieval document.
LLM/RAG Optimization: The schema is explicitly designed to be "self-describing." It includes column definitions and unit annotations to allow LLMs to generate code for plotting, mass modeling, and corrections without external documentation.
Usage Examples: The paper provides three Python examples (using only standard libraries) demonstrating:
1. Multi-component baryonic analysis (SPARC).
2. Handling Tier 2 data with caution zones (WALLABY).
3. Corpus-level parameter-space exploration (histograms and scatter plots).

4. Results

Data Coverage: The corpus spans a wide range of galaxy types and scales, from $\sim3$ kpc dwarf irregulars to $\sim100$ kpc massive spirals, covering peak rotation velocities from low-velocity dwarfs to high-velocity massive systems.
LLM Validation: The author tested four LLMs (Gemini Pro, Claude, AstroSage-Llama-3.1-70B, Copilot Pro) using the per-galaxy JSON files. All models successfully generated syntactically correct Python scripts to:
- Plot rotation curves with error bars.
- Compute baryonic velocities via sign-preserving quadrature.
- Apply empirical omega corrections.
- Result: The schema proved sufficiently self-describing for automated scientific analysis without additional prompting.
Quality Assessment: The tier system successfully distinguishes between high-precision hand-curated data and automated pipeline outputs, allowing users to filter by data provenance.

5. Significance

Reproducibility: By unifying disparate sources into a single schema with consistent units and verified metadata, the corpus removes the "data fragmentation" barrier, enabling reproducible computational astrophysics.
AI Readiness: This is one of the first astrophysical datasets explicitly engineered for Retrieval-Augmented Generation (RAG). It bridges the gap between raw observational data and AI-driven scientific discovery, allowing LLMs to act as analytical agents.
Accessibility: The data is publicly available under CC BY 4.0 on Zenodo, lowering the entry barrier for researchers to perform meta-analyses on galaxy rotation curves.
Future-Proofing: The modular architecture and rigorous verification procedures provide a template for standardizing future HI surveys and preserving data from aging infrastructure.

Known Limitations

The author transparently notes five limitations:

15 THINGS galaxies lack per-point rotation data (metadata only).
SPARC entries lack coordinates and systemic velocities (must be retrieved from NED/SIMBAD).
WALLABY data lacks per-ring uncertainties and baryonic decomposition; velocities $<50$ km/s are unreliable due to beam smearing.
JSON schema keys differ slightly between surveys (data vs. rotation_curve) to reflect genuine observational differences.
Floating-point precision in derived columns may exceed instrumental accuracy; users are advised to round to appropriate precision for uncertainty propagation.