🔭 astrophysics

Data-driven modeling of rotation curves with artificial neural networks

This study employs artificial neural networks to create data-driven models of spiral galaxy rotation curves that capture complex structural features without predefined assumptions, revealing discrepancies with standard theoretical models like Navarro-Frenk-White and highlighting the need for refined physical descriptions of galactic dynamics.

Original authors: Gabriela Garcia-Arroyo, Isidro Gómez-Vargas, J. Alberto Vázquez

Published 2026-07-15

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Gabriela Garcia-Arroyo, Isidro Gómez-Vargas, J. Alberto Vázquez

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, swirling dance floor where stars, gas, and dust spin around a central point. For decades, astronomers have been trying to figure out the "music" that keeps this dance going. They measure how fast the stars are moving at different distances from the center of a galaxy, creating a graph called a "rotation curve." Here's the mystery: when they calculate how much mass (stuff) is visible in the galaxy based on how bright the stars are, it simply isn't enough to explain why the stars are moving so fast. It's like seeing a car zooming down a highway but finding no engine under the hood. To solve this, scientists proposed that invisible "Dark Matter" acts as a hidden engine, providing the extra gravity needed to hold the galaxy together.

Traditionally, to map this invisible engine, scientists use mathematical formulas—like trying to fit a specific shape of cookie cutter onto a lump of dough. They assume the dough (the galaxy) has a certain shape and the cookie cutter (the math model) must fit it perfectly. But galaxies are messy, lumpy, and unique; sometimes the dough doesn't fit the cookie cutter at all. This is where a new, digital approach comes in. Instead of forcing a pre-made shape onto the data, what if we let a computer "learn" the shape of the dough directly from the observations, without guessing what the cookie cutter should look like? This is the question a team of researchers from Mexico and Switzerland set out to answer.

The Digital Detective and the Cosmic Dance

In this study, the researchers acted like digital detectives, but instead of solving a crime, they were trying to understand the hidden structure of 17 different spiral galaxies. They used a powerful type of artificial intelligence called an "Artificial Neural Network" (ANN). You can think of an ANN as a super-smart, digital apprentice. Unlike a traditional scientist who starts with a rulebook (a mathematical formula) and tries to fit the data to it, the ANN starts with nothing but the raw data points—the actual measurements of how fast stars are moving at different distances.

The team fed their AI a high-quality dataset from the "THINGS" survey, which contains precise observations of these 17 specific galaxies. They taught the AI to look at the distance from the center of the galaxy and predict the speed of the stars, along with how confident it was in that prediction. To make sure the AI didn't just memorize the data like a parrot (a problem called "overfitting"), they used a clever trick called "Monte Carlo Dropout." Imagine the AI taking a test multiple times, but every time it takes the test, it randomly forgets a few facts. By seeing how much its answers change when it forgets things, the AI can tell you, "I'm pretty sure about this answer, but I'm a bit shaky about that one." This gives the model a built-in "uncertainty meter."

The Great Showdown: Old Rules vs. New Learning

To see if their new AI method was any good, the researchers pitted it against the old-school method. The traditional method uses a specific mathematical shape for the invisible Dark Matter, called the "Navarro-Frenk-White" (NFW) model. They tested this old model in two ways: first, by assuming the stars' mass was fixed and known (like a strict recipe), and second, by letting the mass of the stars be a "free" variable that could change to make the math work better.

The results were a fascinating mix of success and surprise. The researchers grouped the 17 galaxies into three categories based on which method did the best job of matching the real observations:

The "Old Rules Work Best" Group (Group A): For four galaxies (DDO154, NGC2903, NGC3621, and NGC2403), the traditional mathematical models actually fit the data better than the AI. In these cases, the galaxies seemed to follow the standard "recipe" quite well, and the AI's attempt to learn from the data alone ended up being a bit too wobbly.
The "Middle Ground" Group (Group B): For four other galaxies (NGC3031, NGC2841, NGC3198, and NGC5055), the AI did better than the strict "fixed recipe" model but wasn't quite as good as the "free variable" model. It seems these galaxies were a bit tricky, sitting somewhere between a simple shape and a complex mess.
The "AI Wins" Group (Group C): This was the most exciting group, containing nine galaxies (including NGC3521, NGC2366, and IC2574). For these galaxies, the AI's data-driven model was the clear winner, matching the observations much more closely than any of the traditional mathematical models.

Why the AI Won in Some Cases

The paper suggests that the AI succeeded in Group C because these galaxies have messy, complex structures that don't fit the neat, smooth curves of the old mathematical formulas. For example, some of these galaxies have lumpy distributions of stars or gas, or they might have "clumps" of dark matter that the old models couldn't account for. The traditional models often had to force the data to fit a smooth curve, which led to big errors, especially in the inner parts of the galaxy where the stars are dense. The AI, however, didn't care about smoothness; it just learned the actual, jagged, real-world pattern of the data.

The study highlights that while the old mathematical models are still useful for some galaxies, they have limits. When a galaxy is too complex or has unusual features, the "cookie cutter" approach fails. The AI approach, by contrast, is flexible enough to capture these irregularities without needing to guess the rules beforehand.

What This Means for the Future

The researchers aren't claiming that the old models are wrong or that the AI has solved the mystery of dark matter. Instead, they suggest that we need both tools. The AI acts as a powerful new lens that can show us where the old models break down. It reveals that for certain galaxies, the transition between the visible stars and the invisible dark matter isn't a smooth, predictable slide but a complex, bumpy ride.

By using these data-driven models, scientists can now spot galaxies that behave differently than expected. This doesn't tell us exactly what dark matter is, but it tells us exactly where our current theories might be missing the mark. It's like having a map that not only shows the roads we know but also highlights the rough, unpaved paths where we need to build new bridges. The study concludes that machine learning is a promising partner for astronomers, offering a way to see the universe's complexity without being blinded by our own pre-made assumptions.

Technical Summary: Data-driven modeling of rotation curves with artificial neural networks

Problem Statement
Galactic rotation curves (RCs) are fundamental for determining mass distributions in galaxies. However, a persistent discrepancy exists between the mass inferred from luminosity and the mass derived from observed rotational velocities, a phenomenon typically attributed to Dark Matter (DM). While traditional parametric models (e.g., Navarro-Frenk-White or NFW halos combined with stellar mass-to-light ratios) provide valuable insights, they struggle to capture complex galactic features such as prominent bulges, non-circular motions, and irregular velocity structures. These models rely on simplifying assumptions about mass distribution and specific functional forms, which may limit their ability to reproduce the full dynamical complexity observed in galaxies, particularly in Low-Surface-Brightness (LSB) and High-Surface-Brightness (HSB) systems with distinct gas and stellar contributions.

Methodology
The authors propose a data-driven approach using Artificial Neural Networks (ANNs) to reconstruct rotation curves without relying on predefined astrophysical assumptions or functional forms. The study utilizes high-quality observational data from 17 spiral galaxies in The HI Nearby Galaxy Survey (THINGS).

The technical framework integrates three key components:

Feed-forward Neural Networks: The network takes radial distance as input and outputs two coupled variables: rotational velocity $V(r)$ and its associated statistical uncertainty. This allows the model to learn the global functional relationship $V(r)$ directly from the data.
Monte Carlo Dropout (MC-DO): Used as a probabilistic regularization method. By keeping dropout active during inference and performing multiple stochastic forward passes, the model generates a predictive distribution. This provides an estimate of epistemic uncertainty (model confidence) and mitigates overfitting, which is crucial given the relatively small number of data points (typically hundreds) per galaxy.
Genetic Algorithms (GAs): Employed to optimize the network's hyperparameters (number of hidden layers, neurons per layer, batch size) for each individual galaxy. This global search ensures the selected architecture balances model complexity with generalization capability, avoiding both underfitting and overfitting.

The ANN models are compared against two theoretical parametric models based on the NFW DM profile:

Fixed $\Upsilon_\star$ : Assumes a fixed stellar mass-to-light ratio derived from 3.6 $\mu$ m observations (diet Salpeter IMF).
Free $\Upsilon_\star$ : Treats the stellar mass-to-light ratio as a free parameter to fit the total rotation curve.

Performance is evaluated using the Mean Squared Error (MSE) between the model predictions and observational data.

Key Contributions and Results
The study successfully generates tailored, data-driven models for 17 galaxies, demonstrating that ANNs can capture intricate rotation curve structures without physical priors. The results, analyzed via MSE, lead to a classification of galaxies into three distinct groups:

Group A (Parametric Superiority): Includes DDO154, NGC2903, NGC3621, and NGC2403. In these systems, the parametric models (both fixed and free $\Upsilon_\star$ ) outperform the ANN. The authors note that for DDO154, the observational $\Upsilon_\star$ is particularly precise, and the parametric fit is robust.
Group B (Intermediate Performance): Includes NGC3031, NGC2841, NGC3198, and NGC5055. Here, the ANN performs better than the fixed $\Upsilon_\star$ model but worse than the free $\Upsilon_\star$ model. For NGC5055, the parametric models struggle with the decomposition of two exponential disks, whereas the ANN captures the global trend without resorting to unphysical stellar contributions or excessively massive halos.
Group C (ANN Superiority): Includes nine galaxies (e.g., NGC3521, NGC2366, IC2574). In these cases, the non-parametric ANN reconstruction yields a lower MSE than both parametric models. The authors attribute this to irregular variations in mass distribution, such as clumps of dark matter or stellar substructures, which are difficult to model with smooth parametric profiles. Notably, dwarf LSB galaxies like NGC2366 and IC2574 show the lowest MSEs for the ANN, suggesting that parametric models may fail when the stellar contribution is low or when the mass-to-light ratio assumptions lead to non-physical results (e.g., negative mass).

Significance and Claims
The paper claims that this work highlights the potential of machine learning techniques to identify galaxies whose dynamics are not well captured by standard theoretical models. By avoiding predefined functional forms, the ANN approach reveals where standard parametric profiles (like NFW) break down, particularly in systems with complex mass distributions or structural features like prominent bulges.

The authors emphasize that their method offers a complementary perspective to traditional astrophysics. It does not replace physical models but rather provides a robust, data-driven baseline that can:

Probe the interplay between baryonic and dark matter distributions without assuming specific mass profiles.
Identify limitations in current theoretical descriptions, pointing to the need for more refined physical models.
Handle sparse or noisy data while providing meaningful uncertainty estimates through MC-DO.

The study concludes that while parametric models remain useful for simpler systems, data-driven methods are essential for understanding the full dynamical complexity of galaxies, particularly those where universal profiles fail to describe the transition between baryon and halo domination.

Technical Summary: Data-driven modeling of rotation curves with artificial neural networks

More like this