🔭 astrophysics

MARS: A Lightweight Morphology-Aware RFI Segmentation Network for Mask-Guided Mitigation in Radio Astronomy

The paper introduces MARS, a lightweight, GPU-accelerated RFI mitigation pipeline that utilizes a morphology-aware U-Net to achieve real-time, high-precision radio frequency interference segmentation while effectively preserving dispersed astrophysical signals and significantly outperforming traditional CPU-based methods in speed.

Original authors: Zhaocheng Gong, Jack White, Jayanta Roy, Wesley Armour

Published 2026-08-07

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhaocheng Gong, Jack White, Jayanta Roy, Wesley Armour

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 is a giant, cosmic radio station, broadcasting secrets from pulsars, black holes, and mysterious fast radio bursts. To hear these whispers, astronomers use massive radio telescopes that act like giant ears, listening to the sky. But there's a problem: the sky is noisy. Just like trying to hear a friend at a rock concert, these telescopes are constantly bombarded by "Radio Frequency Interference" (RFI). This isn't just static; it's man-made noise from satellites, cell phones, and microwaves that looks exactly like the cosmic signals astronomers are hunting for. If they don't clean this noise out, they might mistake a microwave oven for a new star, or worse, accidentally delete a real cosmic signal while trying to clean up the mess.

The challenge is speed. Modern telescopes generate data so fast that storing everything to clean it later is impossible. The cleaning has to happen in real-time, right as the data flows in. Traditionally, this cleaning was done by slow, old-school computer programs that couldn't keep up with the speed of modern, high-powered graphics cards (GPUs) used for the rest of the analysis. It was like trying to use a bicycle to catch a speeding train. The question scientists faced was: Can we build a smart, fast, and gentle cleaning system that runs on these powerful GPUs, removes the noise perfectly, but never accidentally throws away the precious cosmic signals hidden inside?

Enter MARS, a new tool designed to solve this exact problem. MARS stands for a "Morphology-Aware RFI Segmentation Network," but you can think of it as a super-smart, shape-sensing digital filter. Instead of just blindly deleting anything that looks weird, MARS is a lightweight neural network (a type of artificial intelligence) trained to recognize the specific "shapes" of human-made noise in the data. It looks at the data as a picture where the horizontal axis is time and the vertical axis is frequency. Human-made noise often looks like long, thin horizontal lines (narrowband signals) or vertical streaks (broadband bursts). MARS is built with special "eyes" that are very good at spotting these specific shapes.

What makes MARS special is that it doesn't just guess; it learns to be careful. The researchers taught MARS to recognize that some cosmic signals, like Fast Radio Bursts (FRBs), can look a bit like noise because they are bright and stretched out. To prevent MARS from accidentally deleting these real signals, they gave it a special rule during training: "If you see a bright, stretched-out signal that looks like a cosmic pulse, don't flag it as noise." This is like teaching a security guard to recognize that a celebrity's red carpet walk looks different from a thief's sneaking, even if both involve moving quickly.

The paper tested MARS in a few different ways. First, they created millions of fake data patches with known noise and known signals to see how well MARS could spot the bad stuff without hurting the good stuff. In these tests, MARS was incredibly accurate, correctly identifying noise 97.8% of the time (an F1 score of 0.978) and being very rarely wrong about what was noise (99.5% precision). More importantly, it kept 97.6% of the fake cosmic signals safe, whereas an older AI method (called RFDL) accidentally deleted nearly half of them.

Next, they tested MARS on a full stream of data, comparing it to the current industry standard tool called filtool. While filtool is a reliable, traditional method, it runs on slower computer processors (CPUs). MARS runs on a powerful GPU. The results showed that MARS cleaned the data just as well as filtool, recovering the same cosmic signals with almost identical strength. However, MARS was much, much faster. On a specific high-end computer chip (the NVIDIA GH200), MARS finished the job 6.2 to 7.0 times faster than the fastest version of filtool running on a top-tier CPU.

The researchers also tested MARS on real data from the Giant Metrewave Radio Telescope (GMRT) in India, looking for two known pulsars. MARS found these pulsars just as clearly as the traditional method did, proving it works on real-world data, not just simulations.

The paper concludes that MARS is a successful, practical solution. It suggests that by using a lightweight AI model that understands the "shape" of noise and is explicitly trained to protect cosmic signals, astronomers can clean their data in real-time without slowing down their search for the universe's secrets. While the authors note that more testing on different telescopes and longer data streams is needed, the current results suggest that MARS is a fast, accurate, and safe way to keep the cosmic radio station clear of static.

Technical Summary: MARS – A Lightweight Morphology-Aware RFI Segmentation Network

Problem Statement
Next-generation radio telescopes generate filterbank data at rates that render storing all observations for offline mitigation impractical. Consequently, Radio Frequency Interference (RFI) mitigation must operate in real or near-real time within search pipelines while preserving dispersed astrophysical signals. Current CPU-based tools (e.g., filtool) create bottlenecks in GPU-centric search pipelines due to data transfer overhead and computational latency. Conversely, existing neural network alternatives are often computationally heavy or lack deployment constraints suitable for high-throughput pipelines. Furthermore, a critical challenge remains: distinguishing between RFI and bright, dispersed astrophysical transients (such as Fast Radio Bursts or single pulses), which can visually resemble interference in the frequency-time plane.

Methodology
The authors present MARS (Morphology-Aware RFI Segmentation Network), a GPU-based mitigation pipeline centered on a lightweight, full-resolution U-Net architecture.

Network Architecture: The model is a reduced-width U-Net (channel widths: 8, 16, 32, 64) designed for TensorRT deployment. It employs a shape-aware bottleneck containing three parallel convolutional branches to capture specific RFI morphologies:
- A local $3 \times 3$ branch for compact or irregular interference.
- A horizontal $1 \times 9$ branch for narrowband structures extended in time.
- A vertical $9 \times 1$ branch for broadband or short-duration structures extended in frequency.
- Decoder refinement blocks utilize anisotropic residual modules ( $1 \times 31$ and $31 \times 1$ convolutions) to expand the receptive field without increasing channel width.
Pipeline Workflow: MARS does not directly reconstruct cleaned data. Instead, it predicts a soft RFI mask which is applied to the original data via an explicit replacement pipeline. The workflow includes:
1. Preprocessing: Per-channel robust normalization (median and scaled median absolute deviation) and bounded $\tanh$ compression.
2. Inference: Processing fixed $512 \times 512$ patches on the GPU.
3. Mask Application: Replacing contaminated samples in a mean/std-normalized stream, followed by baseline removal and rescaling to 8-bit filterbank format.
4. Hardware: The entire pipeline, from normalization to output rescaling, is implemented on the GPU to avoid CPU-GPU data transfers. Inference is accelerated using TensorRT (FP16 precision).
Training Strategy: The model is trained on synthetic frequency-time patches with known RFI masks. A key innovation is the inclusion of an astronomical-signal preservation loss ( $L_{astro}$ ). This term penalizes high-confidence RFI predictions on pixels containing injected synthetic dispersed pulses (simulating FRBs or single pulses), specifically targeting low-Dispersion Measure (DM) and high-Signal-to-Noise (S/N) cases where pulses resemble interference.

Key Contributions

Morphology-Aware Design: The explicit use of local, horizontal, and vertical filters allows the network to capture anisotropic RFI structures without requiring a large, generic segmentation backbone.
Signal Preservation Mechanism: The introduction of $L_{astro}$ significantly reduces the false masking of dispersed astrophysical signals, a common failure mode in RFI detection.
GPU-Native Pipeline: Unlike previous neural approaches that may require CPU intervention, MARS implements the full mitigation cycle (normalization, inference, replacement, rescaling) on the GPU, enabling seamless integration with GPU-accelerated search pipelines like AstroAccelerate.
Operational Baseline: The system produces standard cleaned filterbank files compatible with established tools like PRESTO and SIGPROC, facilitating direct comparison with traditional methods.

Results

Patch-Level Performance: On a synthetic benchmark of 15,000 patches across six RFI families, MARS achieved an F1 score of 0.978 and precision of 0.995, outperforming the RFDL neural baseline (F1: 0.924, Precision: 0.903). MARS showed particular robustness in narrowband RFI detection where RFDL suffered from low precision.
Signal Protection: In tests involving injected dispersed pulses, MARS retained 97.6% of fluence in clean patches and 96.4% in patches with mixed RFI. In contrast, the RFDL baseline retained only ~57% and ~53%, respectively. Ablation studies confirmed that removing $L_{astro}$ reduced retention to ~88% and ~83%, highlighting the loss term's critical role in protecting low-DM, high-S/N pulses.
Filterbank-Level Recovery: When applied to synthetic binary-pulsar filterbanks and searched with PRESTO, the GPU pipeline recovered target candidates with significance ratios of 0.90–0.99 relative to filtool. In two real GMRT observations (J0139+5814 and J0659+1414), MARS recovered known pulsars with significance comparable to or slightly higher than filtool.
Computational Efficiency: On an NVIDIA GH200 GPU, MARS achieved a 6.2×–7.0× speedup over the fastest multi-threaded filtool configurations on an AMD EPYC 9825 CPU. Notably, the full MARS pipeline (including replacement and rescaling) was 69–73× faster than the RFDL baseline running on the same GPU for mask prediction alone.

Significance and Claims
The paper claims that MARS provides a practical, GPU-accelerated alternative for RFI mitigation in pulsar-search workflows. It addresses the dual constraints of computational speed and scientific fidelity by:

Enabling Real-Time Mitigation: By keeping data on the GPU and utilizing a lightweight model, it removes the CPU bottleneck inherent in traditional pipelines.
Balancing Removal and Preservation: It demonstrates that high-precision RFI masking can be achieved without sacrificing dispersed astrophysical signals, provided specific loss functions and morphology-aware architectures are employed.
Operational Compatibility: The method produces standard output formats, allowing it to function as a drop-in replacement for CPU-based tools in existing search pipelines without requiring changes to downstream analysis software.

The authors modestly note limitations, including the reliance on controlled synthetic injections for quantitative benchmarks and the need for broader validation across diverse telescopes and RFI environments. However, the results suggest that the proposed approach successfully bridges the gap between high-throughput processing requirements and the rigorous signal preservation needed for transient searches.

Technical Summary: MARS – A Lightweight Morphology-Aware RFI Segmentation Network

More like this