🔭 astrophysics

Deep Learning-Based Classification and Analysis of Pulsar Candidates in Fermi-LAT Unassociated Sources

This paper introduces a hierarchical 1D-CNN framework called TabularResCNN that classifies unassociated Fermi-LAT sources by analyzing their spectral shapes and variability without relying on spatial coordinates, successfully identifying 202 high-confidence pulsar candidates (including 5 confirmed by FAST) and increasing the known pulsar population by over 60%.

Original authors: C. Pozo González, R. López-Coto, J. Méndez-Gallego, E. de Oña Wilhelm

Published 2026-08-03

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: C. Pozo González, R. López-Coto, J. Méndez-Gallego, E. de Oña Wilhelm

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, cosmic radio station that is constantly broadcasting on many different frequencies. Some of these signals are loud and clear, like a pop song on the radio, while others are faint whispers lost in the static. For decades, astronomers have been using a powerful space telescope called Fermi-LAT to listen to the high-energy "gamma-ray" part of this cosmic broadcast. They have mapped out thousands of these sources, but a huge chunk of them—about one-third—remain a mystery. These are the "unassociated" sources. They are like radio stations that are clearly broadcasting, but no one knows who is singing the song or what kind of music they play. Are they ancient, spinning neutron stars (pulsars) that have been recycled and sped up? Are they young, energetic stars just born from supernova explosions? Or are they distant, super-massive black holes (AGNs) feasting on gas?

To solve this mystery, scientists have traditionally tried to guess the identity of these sources by looking at where they are in the sky. It's a bit like trying to guess someone's accent just by looking at a map of where they live. While this works okay, it's not perfect and can lead to mistakes, especially for sources that don't fit the usual patterns. This is where a new kind of detective work comes in: Artificial Intelligence. Instead of just looking at the map, these new AI models listen to the "shape" of the signal itself. They analyze how the energy of the light changes across different bands, looking for the unique fingerprint of the source. The goal is to find the hidden pulsars among the crowd, which is crucial because pulsars are cosmic laboratories that help us understand gravity, time, and the extreme physics of the universe.

The Cosmic Detective: Teaching AI to Listen to the Stars

In this paper, a team of astronomers from Spain and Germany decided to build a super-smart digital detective to solve the mystery of these unassociated gamma-ray sources. They didn't just want to guess; they wanted to teach a computer to "see" the shape of the light in a way that humans can't easily do.

The Old Way vs. The New Way
Imagine you are trying to identify different fruits in a dark room. The old method was like asking, "Is this fruit in the basket on the left or the right?" If you knew that apples usually sit on the left and oranges on the right, you could guess correctly most of the time. But if you found a fruit in the middle, you'd be stuck. This is what older computer programs did: they looked at the location of the star in the sky to guess what it was.

The new method, used in this paper, is like turning on a light and looking at the fruit's skin, texture, and smell. The researchers built a special type of AI called a 1D Convolutional Neural Network (or 1D-CNN for short). Think of this AI as a very curious teenager with a magnifying glass. Instead of just looking at a list of numbers, this AI slides its "magnifying glass" across the data, looking for patterns and shapes in the way the light energy flows. It treats the data like a continuous song rather than a list of separate notes. This allows it to spot the "curves" and "breaks" in the light that are unique to pulsars, ignoring the location entirely.

The Two-Step Sorting Game
The team created a two-stage game for their AI to play:

Stage One: The AI looks at a source and asks, "Is this a black hole monster (AGN) or a spinning neutron star (Pulsar)?" It separates the crowd into two big piles.
Stage Two: For the pile of spinning stars, the AI asks, "Is this a young, energetic star (Young Pulsar) or an old, recycled one (Millisecond Pulsar)?"

Why They Didn't Fake the Data
Usually, when computer scientists train an AI, they might create fake examples to help the computer learn rare things. Imagine if you had only 5 pictures of tigers and 100 pictures of cats, so you made up 95 fake tiger pictures to balance the books. The researchers in this paper said, "No thanks!" They argued that making up fake data for these cosmic sources could trick the AI into learning the wrong patterns. Instead, they taught the AI to pay extra attention to the few real pulsars it actually saw, making sure it didn't ignore them just because they were rare.

The Big Findings
The team fed their AI the data for 2,563 mysterious, unassociated sources from the Fermi-LAT catalog. Here is what the AI found:

It identified 1,136 sources as black holes (AGNs).
It found 202 high-confidence candidates for pulsars.
- 166 of these were Young Pulsars.
- 36 were Millisecond Pulsars.

This is a huge deal because it increases the known population of gamma-ray pulsars by more than 60%.

Does the AI Know What It's Talking About?
To make sure the AI wasn't just guessing, the researchers checked if their new candidates made sense with what we already know about the universe:

Location Check: Even though the AI never looked at the location of the stars while learning, the Young Pulsars it found were all huddled tightly around the flat disk of our galaxy (the Milky Way), just like real young stars should be. The older Millisecond Pulsars were spread out more widely, which is exactly what we expect for older stars that have been kicked around by gravity over time.
The "FAST" Test: The ultimate test came from a giant radio telescope on Earth called FAST. Recently, FAST discovered 5 new pulsars by looking for them blindly. The researchers checked their AI's list and found that all 5 of those new pulsars were already on their list as high-confidence candidates! This means the AI successfully predicted these discoveries before the radio telescope even confirmed them.

How the AI "Thinks"
The researchers also used a special tool called Grad-CAM to see which parts of the data the AI was focusing on. It turned out the AI was using real physics to make its decisions:

To spot black holes, it looked at the low-energy light and how much the signal flickered (black holes are often messy and flickery).
To spot pulsars, it looked at the high-energy light and the stable, smooth curves of the signal.
To tell young pulsars from old ones, it looked at the very highest energy bands and how uncertain the measurements were.

The Bottom Line
This paper shows that by teaching AI to listen to the "shape" of the light rather than just looking at a map, we can find hidden cosmic treasures. The team has provided a "hit list" of 202 new pulsar candidates that are ready for radio telescopes like FAST, MeerKAT, and the future Square Kilometre Array (SKAO) to investigate. It's like giving astronomers a treasure map that points directly to the gold, saving them time and helping them discover the next generation of cosmic wonders.

Technical Summary: Deep Learning-Based Classification and Analysis of Pulsar Candidates in Fermi-LAT Unassociated Sources

Problem Statement
The Fourth Fermi-LAT Source Catalog (4FGL-DR4) contains approximately one-third of its sources as "unassociated," lacking firm multi-wavelength counterparts. Properly classifying these objects is essential for understanding the demographics of Active Galactic Nuclei (AGNs) and Pulsars (PSRs). While machine learning (ML) has become the standard for this task, existing approaches face two primary limitations:

Spatial Bias: Many models rely heavily on Galactic coordinates (e.g., latitude) as training features. While this yields high accuracy by exploiting the known concentration of Young Pulsars (YPs) in the Galactic disk, it risks overfitting to the telescope's selection function and fails to detect outlier populations (e.g., high-latitude Millisecond Pulsars) that do not conform to spatial priors.
Loss of Topological Information: Traditional tree-based models (e.g., Random Forest, XGBoost) treat spectral and temporal features as independent tabular entries. This approach ignores the sequential order and physical correlations inherent in Spectral Energy Distributions (SEDs), such as spectral curvature and cutoffs, which are critical for distinguishing source types.

Methodology
The authors propose a hierarchical deep learning framework centered on a custom 1D Convolutional Neural Network (CNN) named TabularResCNN. The methodology is designed to exploit the intrinsic topological structure of the data without relying on spatial coordinates.

Data Preprocessing & Input Strategy:
- Feature Engineering: The model utilizes 79 features from the 4FGL-DR4 catalog, including spectral parameters, variability indices, and detection significance across 7 energy bands.
- Array Expansion: Multi-dimensional FITS columns (e.g., flux across energy bands) are flattened and sequentially arranged to preserve energy-dependent ordering, allowing the CNN to interpret the SED as a coherent 1D signal.
- Dual-Channel Input: To handle missing values and upper limits without destroying information via standard imputation, the model uses two input channels: a "Value Map" (normalized numerical features) and a "Sensitivity Mask" (binary tensor indicating valid measurements vs. missing data). This allows the network to learn patterns of observability.
Hierarchical Classification Framework:
- Stage 1 (AGNs vs. PSRs): A binary classifier separates sources into AGNs and Pulsars. This stage leverages the dichotomy where AGNs typically exhibit high variability and softer spectra, while PSRs show stable, curved spectra with GeV cutoffs.
- Stage 2 (YPs vs. MSPs): Sources classified as PSRs are further categorized into Young Pulsars (YPs) and Millisecond Pulsars (MSPs).
Model Architecture:
- The TabularResCNN employs 1D convolutional layers to scan local neighborhoods for gradients (e.g., spectral breaks) and inter-band correlations.
- Residual blocks with skip connections facilitate deep hierarchical learning without degradation.
- Global Average Pooling (GAP) reduces parameter counts and mitigates overfitting.
Training Strategy & Imbalance Mitigation:
- Cost-Sensitive Learning: Instead of synthetic oversampling (e.g., SMOTE), which the authors argue introduces statistical artifacts and covariate shifts by reinforcing biases toward bright sources, they employ a Weighted Cross-Entropy Loss. This penalizes the misclassification of minority classes (PSRs) inversely proportional to their frequency.
- Threshold Optimization: Rather than using a default 0.5 probability cutoff, the authors implement a cost-sensitive threshold optimization based on "observational efficiency curves." This shifts the paradigm from mathematical accuracy to the operational reality of radio follow-up, prioritizing high-purity candidate lists over absolute completeness.
- Interpretability: Grad-CAM is used to visualize which spectral features drive the model's decisions, ensuring physical interpretability.

Key Results
Applied to 2,563 unassociated sources in the 4FGL-DR4 catalog, the framework yielded the following results:

Classification Performance:
- Stage 1: Achieved 96.7% overall accuracy in separating AGNs from PSRs, with near-perfect retrieval for AGNs and high precision for PSRs.
- Stage 2: Achieved 81.4% accuracy in distinguishing YPs from MSPs, a more challenging task due to spectral overlap.
Candidate Identification:
- Identified 1,136 high-confidence AGN candidates and 202 high-confidence PSR candidates (166 YPs and 36 MSPs).
- This represents a >60% increase in the known population of Fermi-LAT pulsars.
- By applying cost-sensitive thresholds, the authors filtered out 847 low-confidence candidates that would have been included under a standard 0.5 cutoff, resulting in a highly robust catalog.
Physical Consistency:
- Spatial Distribution: Despite not using spatial coordinates in training, the predicted YPs were strictly confined to the Galactic plane ( $|b| < 5^\circ$ ), while MSPs showed a broader vertical distribution, and AGNs were isotropic.
- P-Ṗ Diagram: Cross-matching with the ATNF catalog confirmed that classified YPs and MSPs occupy the correct regions of the spin-period vs. period-derivative diagram, validating that the model learned intrinsic magnetospheric physics.
- External Validation: The model correctly classified 5 out of 5 pulsars recently confirmed by the Five-hundred-meter Aperture Spherical Radio Telescope (FAST) as high-confidence candidates prior to their radio confirmation.

Significance and Claims
The paper claims that the proposed framework successfully decouples intrinsic emission physics from line-of-sight spatial biases, a limitation prevalent in previous studies. By utilizing the sequential topology of spectral data via 1D-CNNs, the model identifies pulsar candidates based on their physical emission properties rather than their location in the sky.

The authors emphasize that their approach avoids the methodological risks of synthetic oversampling, which can distort the physical reality of faint sources. The resulting catalog provides a "bias-minimized" target list for deep radio surveys. The 100% agreement with recent blind-search discoveries by FAST serves as empirical validation of the model's predictive power. The authors conclude that this curated list offers a prioritized roadmap for current and future facilities, including FAST, MeerKAT, the Square Kilometre Array Observatory (SKAO), and the Cherenkov Telescope Array Observatory (CTAO), to maximize discovery yields in the search for new pulsars.

The Cosmic Detective: Teaching AI to Listen to the Stars

More like this