🔭 astrophysics

$\mbox{H}$ $\mbox{I}$ 21-cm Absorption Spectra Classification using Machine Learning

This paper demonstrates that a Random Forest machine learning model, trained on spectral parameters of H I 21-cm absorption lines, can efficiently and accurately classify them as either "intervening" or "associated" with an accuracy of 89%, offering a scalable solution for future large-scale surveys like FLASH and the Square Kilometre Array.

Original authors: Debasish Mondal, Anirudh S. Nemmani, Arunima Banerjee

Published 2026-08-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Debasish Mondal, Anirudh S. Nemmani, Arunima Banerjee

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

The universe is filled with vast clouds of cold, atomic hydrogen gas, the fundamental building block from which stars and galaxies are born. Astronomers have long known how to find this gas when it is close to us and glowing brightly, but detecting it in the distant, cold, and dark reaches of space is far more difficult. To find these invisible reservoirs, scientists often look for a specific shadow: when a bright, distant radio source, like a quasar, shines through a cloud of hydrogen, the gas absorbs a tiny fraction of the light at a very specific frequency. This creates a dip, or an absorption line, in the radio signal. By studying these dips, researchers can learn about the temperature, motion, and location of the gas. However, a critical question remains: is this gas sitting right next to the bright radio source, or is it a separate galaxy lying in the foreground, blocking the view? Distinguishing between these two scenarios usually requires expensive, time-consuming follow-up observations with optical telescopes to measure the exact distance of the radio source. With new, massive surveys on the horizon that promise to find thousands of these gas clouds, astronomers need a faster way to sort them out without waiting for years of follow-up data.

A team of researchers has developed a new method to solve this sorting problem using machine learning, a type of computer program that learns to recognize patterns from examples. Instead of relying on slow optical observations, they taught a computer to look directly at the shape of the radio absorption lines themselves. The team gathered a collection of 118 known examples of these gas clouds, carefully measuring the specific shape of each absorption line. They used a sophisticated mathematical tool called the Busy function to describe the curves, which is better at capturing the complex, often uneven shapes of these gas clouds than older, simpler methods. From these shapes, the computer extracted key details, such as how wide the absorption line is and how deep the dip goes. The researchers then trained six different machine learning models to look at these details and decide whether the gas was "associated" with the radio source or "intervening" from a foreground galaxy.

The results showed that one particular model, known as a random forest, was the most effective at making these distinctions. This model achieved a success rate of 89 percent in correctly identifying the type of gas cloud, a significant improvement over previous attempts that relied on simpler curve-fitting methods. The study revealed that the most important clue for the computer was simply the width of the absorption line. Gas clouds that are physically connected to the radio source tend to have much broader and stronger absorption lines, likely due to the violent motion of gas swirling around a central black hole. In contrast, gas in foreground galaxies produces narrower, weaker lines. The researchers found that they did not need the full, complex description of the curve to make a good guess; using just the width of the line and the total depth of the absorption was enough to achieve nearly the same level of accuracy.

To prove that this new approach works for real-world applications, the team applied their best-performing model to 30 newly discovered gas clouds found in a recent survey called FLASH. The model successfully categorized these new discoveries, agreeing with previous expert classifications about 80 percent of the time. This demonstrates that the technique can be used immediately to process the massive flood of data expected from future, large-scale radio telescopes. By automating the classification process, astronomers will be able to quickly separate the local gas from the distant foreground, allowing them to study the distribution and evolution of cold gas across the universe without being bottlenecked by the need for follow-up observations. The work confirms that the shape of a radio signal holds the key to understanding the location of the gas creating it, turning a complex astronomical puzzle into a manageable data task.

Technical Summary: H i 21-cm Absorption Spectra Classification using Machine Learning

Problem Statement
H i 21-cm absorption is a critical tool for studying cold atomic hydrogen gas in galaxies, arising either from intervening galaxies along the line of sight or from the radio source itself (associated). Distinguishing between these two origins is essential for statistical studies of galaxy evolution and Active Galactic Nucleus (AGN) feedback. Traditionally, this classification requires follow-up optical/infrared spectroscopy to determine the redshift of the radio source. However, this approach is often unfeasible for large, blind surveys due to the faintness of sources, the lack of spectral lines in objects like blazars, and the time-consuming nature of obtaining multi-band photometric data for photometric redshift estimation. Previous machine learning (ML) attempts to automate this classification (e.g., Curran et al. 2016; Curran 2021) achieved accuracies of approximately 80% but relied on fitting absorption spectra with Gaussian profiles. The authors note that Gaussian profiles are symmetric and often fail to accurately describe the broad, asymmetric, double-horned profiles characteristic of spiral galaxies, potentially limiting classification performance.

Methodology
The authors developed a supervised ML framework to classify H i 21-cm absorption lines as either "intervening" or "associated" without requiring redshift information.

Data Sample: The study utilized a dataset of 118 known H i 21-cm absorption spectra (74 associated, 44 intervening) collected from the literature.
Spectral Fitting: Instead of Gaussian profiles, the authors employed the Busy function, a versatile analytic function combining error functions and a polynomial, to fit the spectral profiles. This method allows for the modeling of steep flanks and flat troughs typical of double-horned galaxy spectra. From the Busy function fits, 13 spectral parameters were extracted (including amplitude, slopes, offsets, linewidths $w_{50}$ and $w_{20}$ , integrated optical depth $\tau_{int}$ , and signal-to-noise ratio).
Feature Selection: A correlation analysis revealed a strong correlation between $w_{50}$ and $w_{20}$ ( $r=0.95$ ). Consequently, $w_{50}$ was excluded to reduce dimensionality, leaving 12 spectral parameters plus the SNR as predictors.
Machine Learning Models: Six supervised classification algorithms were trained and evaluated: Gaussian Naive Bayes, Logistic Regression, Decision Tree, Random Forest, Support Vector Machine (SVM), and XGBoost.
Training Protocol: To address the small sample size and class imbalance, the authors:
1. Balanced the training set by randomly selecting 44 associated absorbers to match the 44 intervening absorbers.
2. Used a 80:20 train-test split with stratified sampling.
3. Applied standard scaling (Z-score normalization).
4. Performed hyperparameter tuning using Grid Search Cross-Validation combined with Leave-One-Out (LOO) techniques to ensure unbiased estimates.
5. Repeated the entire training and evaluation process 1,000 times to generate robust average performance metrics (Accuracy, $F_1$ -score, AUC).
Comparative Analysis: The authors also tested the models using parameters extracted from multi-Gaussian fits and evaluated the impact of applying a redshift cut ( $z_{abs} \geq 0.1$ ). Finally, they tested a reduced model using only the two most significant parameters: linewidth ( $w_{20}$ ) and integrated optical depth ( $\tau_{int}$ ).

Key Results

Model Performance: The Random Forest model consistently outperformed the other classifiers. When trained on all 12 spectral parameters and SNR, it achieved an average test accuracy of 89%, an $F_1$ -score of 0.90, and an AUC score of 0.94.
Feature Importance: Analysis of the Random Forest model identified the linewidth parameter $w_{20}$ as the most significant predictor, followed by the integrated optical depth ( $\tau_{int}$ ). This aligns with physical expectations where associated absorbers (AGN) typically exhibit broader and stronger lines than intervening absorbers (star-forming galaxies).
Reduced Parameter Set: Retraining the Random Forest model using only $w_{20}$ and $\tau_{int}$ (plus SNR) yielded an accuracy of 88%, an $F_1$ -score of 0.88, and an AUC of 0.91. This suggests that these two parameters are largely sufficient for classification.
Fitting Method Comparison: The Busy function fitting approach yielded classification results comparable to (and slightly better than in the redshift-cut scenario) the multi-Gaussian fitting approach used in previous studies. The Busy function offers the advantage of a fixed number of free parameters (8) regardless of profile complexity, whereas multi-Gaussian fits require estimating 3 parameters per component, leading to variable complexity (1–5 Gaussians, or 3–15 parameters).
Redshift Independence: Applying a redshift cut ( $z_{abs} \geq 0.1$ ) reduced the test accuracy only marginally (from 89% to 87%), indicating that the spectral parameters themselves are robust classifiers independent of redshift evolution effects within the sample.
Application to New Data: The authors applied the retrained Random Forest model (using $w_{20}$ and $\tau_{int}$ ) to 30 new H i 21-cm absorbers detected in the FLASH pilot survey. The model's predictions agreed with the logistic regression classification of Curran (2021) in approximately 80% of cases.

Significance and Claims
The paper claims to provide an efficient, automated ML framework for classifying H i 21-cm absorbers, which is crucial for upcoming large-scale blind surveys with Square Kilometre Array (SKA) precursors (e.g., ASKAP, MeerKAT) and the SKA itself. The primary contributions are:

Methodological Improvement: Demonstrating that the Busy function is a suitable and efficient alternative to Gaussian fitting for extracting spectral parameters for ML classification, particularly for asymmetric profiles.
Performance Enhancement: Achieving higher classification accuracy (89% vs. ~80%) compared to previous studies, attributed to the use of the Busy function and a more comprehensive set of ML algorithms.
Feature Simplification: Establishing that a minimal set of two spectral parameters ( $w_{20}$ and $\tau_{int}$ ) is sufficient for high-accuracy classification, simplifying the data requirements for future surveys.
Scalability: The developed techniques are presented as a viable solution for handling the hundreds to thousands of new absorbers expected in future surveys, reducing the reliance on time-consuming follow-up spectroscopy.

The authors maintain a modest tone, noting that while the current results are promising, the predictive power of these models is expected to increase as larger datasets become available for training in the SKA era.

More like this