FRTSearch: Unified Detection and Parameter Inference of Fast Radio Transients using Instance Segmentation
Authors: Bin Zhang, Yabiao Wang, Xiaoyao Xie, Shanping You, Xuhong Yu, Qiuhua Li, Hongwei Li, Shaowen Du, Chenchen Miao, Dengke Zhou, Jianhua Fang, Jiafu Wu, Pei Wang, Di Li
Organizations: School of Mathematical Sciences, Guizhou Normal University, Guiyang 550001, China · Guizhou Provincial Key Laboratory of Information and Computing Science , Guizhou Normal University, Guiyang 550001, China · School of Cyber Science and Technology, Guizhou Normal University, Guiyang 550001, China · Tencent Youtu Lab, Shanghai 201103, China · College of Physics and Electronic Engineering, Qilu Normal University, Jinan 250200, China · Zhejiang Lab, Hangzhou, Zhejiang 311121, China · National Astronomical Observatories, Chinese Academy of Sciences, Beijing 100101, China · Institute for Frontiers in Astronomy and Astrophysics, Beijing Normal University, Beijing, 102206, China · State Key Laboratory of Radio Astronomy and Technology, Chinese Academy of Sciences, A20 Datun Road, Chaoyang District, Beijing, 100101, China · New Cornerstone Science Laboratory, Department of Astronomy, Tsinghua University, Beijing 100084, China
The exponential growth of data from modern radio telescopes presents a significant challenge to traditional single-pulse search algorithms, which are computationally intensive and prone to high false-positive rates due to Radio Frequency Interference (RFI). In this work, we introduce FRTSearch, an end-to-end framework unifying the detection and physical characterization of Fast Radio Transients (FRTs). Leveraging the morphological universality of dispersive trajectories in time-frequency dynamic spectra, we reframe FRT detection as a pattern recognition problem governed by the cold plasma dispersion relation. To facilitate this, we constructed CRAFTS-FRT, a pixel-level annotated dataset derived from the Commensal Radio Astronomy FAST Survey (CRAFTS), comprising 2{,}392 instances across diverse source classes. This dataset enables the training of a Mask R-CNN model for precise trajectory segmentation. Coupled with our physics-driven IMPIC algorithm, the framework maps the geometric coordinates of segmented trajectories to directly infer the Dispersion Measure (DM) and Time of Arrival (ToA). Benchmarking on the FAST-FREX dataset shows that FRTSearch achieves a 98.0% recall, competitive with exhaustive search methods, while reducing false positives by over 99.9% compared to PRESTO and delivering a processing speedup of up to 13.9×. Furthermore, the framework demonstrates robust cross-facility generalization, detecting all 19 tested FRBs from the ASKAP survey without retraining. By shifting the paradigm from search-then-identify'' to detect-and-infer,'' FRTSearch provides a scalable, high-precision solution for real-time discovery in the era of petabyte-scale radio astronomy.
Figures & tables
Figure 1: Waterfall plots for three typical source classes detected by the CRAFTS survey with the FAST telescope: PSR B1911 − 04 (left), RRAT J2325 − 0530 (middle), and FRB 20121102 (right). Each plot displays a time window beginning at the source’s ToA. Inset Panels: The corresponding de-dispersed pulse profiles, obtained by aligning the signal in time using the source’s DM.
Figure 2: Visualization of key stages in the FRTSearch pipeline using seven representative examples (resized to 256×256 ). Columns (a)–(d): Four distinct FRT sources exhibiting varied dispersive features. Columns (e)–(g): Pulsar B0820 + 02 observed simultaneously by three adjacent FAST receiver beams, demonstrating signal variation. Row 1: Preprocessed time-frequency dynamic spectra (model input). Row 2: Manually annotated ground-truth masks for training. Row 3: Mask R-CNN predictions displaying segmentation masks (pink contours), bounding boxes (pink rectangles), and confidence scores. Row 4: De-dispersed dynamic spectra with IMPIC-inferred DM values.
Figure 3: Distribution of observational parameters (DM, S/N, and pulse width) across the full dataset of 2,392 FRT instances.
No.
Source Name
R.A. (hh:mm)
Decl. (dd:mm)
DM (pc cm -3 )
Files
FRT Instances
Set
1
PSR B0559-05
06:01
− 05:27
80.5
17
278
Training
2
PSR B0621-04
06:24
− 04:24
70.8
12
88
Training
3
PSR B1841-05
18:44
− 05:38
411.7
10
299
Training
4
PSR B1846-06
18:49
− 06:37
148.2
6
15
Training
5
PSR J0458-0505
04:58
− 05:05
47.8
14
28
Training
6
PSR J0459-0210
04:59
− 02:10
21.0
14
77
Training
Table 1: Detailed composition and train-validation splitting of the CRAFTS-FRT dataset.
Figure 4: The end-to-end FRTSearch pipeline automatically processes radio observation files, generating diagnostic plots and a list of associated parameters for FRT candidates.
Figure 5: The architecture of the Mask R-CNN model designed for FRT detection.
Backbone
Optimizer
Base LR
Weight Decay
Specific Settings
ResNet/ResNeXt/HRNet
SGD
0.02
10−4
Momentum = 0.9
Swin-T
AdamW
10−4
0.05
β=(0.9,0.999) , zero decay for pos_embed/norm/bias
ConvNeXt
AdamW
10−4
0.05
β=(0.9,0.999) , layer-wise decay rate = 0.95
Table 2: Optimization configurations for different backbone architectures.
Backbone
TP
# Candidates
Segmentation Metrics
GT-level Metrics
File-level Metrics
bbox AP (%)
segm AP (%)
Recall (%)
FPPI
Recall (%)
FPPI
ResNet-18
1237
2143
75.8
52.7
97.4
1.5
100.0
9.0
ResNet-50
1001
1839
72.9
49.6
92.7
1.4
96.0
8.3
ResNet-101
1058
3046
72.9
52.3
95.9
3.3
97.0
19.8
ResNeXt-101
990
2395
74.7
51.0
95.2
2.3
97.0
14.0
HRNet ⋆
887
1357
77.9
57.3
95.1
0.8
100.0
4.7
Table 3: Backbone architecture comparison on the validation set (100 files, 587 GT instances). The selected model is marked with ⋆ .
Figure 6: Evaluation of IMPIC hyperparameters. The plot compares the mean R2 score (left axis, solid lines) and processing time (right axis, dashed line) across different sample sizes ( Nsample ) and iterations ( Niter ). The red star highlights the adopted configuration ( Nsample=100 , Niter=15 ) used in this work.
Source
Ref DM (pc cm -3 )
TP
Mean DM (pc cm -3 )
Std (pc cm -3 )
R2
1σ (%)
2σ (%)
RRAT J2325-0530
14.9
3
19.7
2.1
0.969
66.7
100.0
PSR B0820+02
23.7
325
30.0
14.9
0.965
92.6
95.1
PSR B1911-04
89.4
103
88.3
3.9
0.991
90.3
97.1
PSR J1820-0509
102.4
109
101.5
4.0
0.995
89.9
96.3
PSR J0652-0142
116.3
42
113.8
6.2
0.992
73.8
95.2
PSR J1948+2333
198.2
72
200.1
14.4
0.995
90.3
95.8
Table 4: DM Inference Statistics for Validation Sources. Columns: reference DM; number of true positive detections (TP); mean and standard deviation of inferred DMs; mean coefficient of determination ( R2 ) of the trajectory fit; and fractional accuracy within ±1σ and ±2σ tolerance.
Figure 7: Error correlation with observational parameters. Scatter plots show absolute ToA residuals (top row: a–c) and DM residuals (bottom row: d–f) versus S/N (left column), pulse width (middle column), and trajectory completeness (right column). Points are color-coded by detection confidence scores.
Figure 8: Diagnostic plots for three representative candidates. Each column displays: (top) de-dispersed pulse profile with on-pulse region (vertical dotted lines) and metadata; (middle) de-dispersed dynamic spectrum with frequency-averaged bandpass (right); (bottom) dispersed dynamic spectrum showing frequency-time trajectory. (a) Candidate using IMPIC-inferred parameters. (b–c) FRB 20121102 (missed by FRTSearch ) using parameters provided by FAST-FREX [ 25 ] .
Method
Threshold
TP
#Candidates
Recall (%)
FPPI (per file)
Time (s/file)
PRESTO
S/N ≥ 5
479
2,990,577
79.8
4,983.5
221.8
Heimdall
S/N ≥ 5
505
29,901
84.2
49.0
16.0
TransientX
S/N ≥ 5
600
49,405
100.0
81.3
60.3
FRTSearch
Confidence ≥ 0.1
588
3,050
98.0
4.1
8.7 / 16.0
Table 5: Benchmark comparison on the FAST-FREX dataset [ 25 ] (600 FRB bursts). Best results are highlighted in bold . For FRTSearch , execution time is reported as parameter-only / with diagnostic plots .
Parameter
CRAFTS
ASKAP
File Format
PSRFITS
Filterbank
Sampling Time ( μ s)
98.3
1266.5
Channel Width (MHz)
0.122
1.0
Bit Depth
2
8
Integration Time (s)
12.88
3345.33
Frequency Range (MHz)
1000–1500
1130–1465
Table 6: Comparison of instrumental parameters between CRAFTS-FRT (training data) and ASKAP (test data).
Figure 9: FRTSearch detection results on 19 ASKAP FRBs [ 62 ] . The panels show time-frequency spectra ( 336×336 pixels) with overlaid model predictions: segmentation masks (pink contours), bounding boxes (pink rectangles), and confidence scores (white text).
Figure 10: Detection efficiency of FRTSearch on the FAST-FREX benchmark. The recall rate is evaluated as a function of S/N, revealing a progressive decrease in detection reliability for S/N<10 .
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
FRB Name
R.A.
Decl.
S/N
Published DM
Inferred DM
Inferred ToA
Confidence
(hh:mm)
(dd:mm)
(pc cm -3 )
(pc cm -3 )
(s)
FRB20170416
22:13
− 55:02
8.3
523.2
494.0
3260.7
0.58
FRB20170428
21:47
− 41:51
8.2
991.7
1070.3
1893.5
0.36
FRB20170707
02:59
− 57:16
10.5
235.2
239.5
570.1
0.82
FRB20170712
22:36
− 60:57
13.8
312.8
310.6
2638.1
0.97
FRB20170906
21:59
− 19:57
18.6
390.3
385.1
2016.1
0.84
Appendix
Table 7: Detection performance of FRTSearch on 19 ASKAP FRBs. Published parameters from [ 62 ] ; inferred DM, ToA, and confidence scores from IMPIC and Mask R-CNN.
Fast Radio Bursts (FRBs) are millisecond-duration radio transients whose automated detection increasingly relies on highly specialized deep learning models. These detectors achieve exceptional performance, but they require large task-specific training datasets and cannot be redefined without retraining. In this work, we evaluate whether small, open-weight, locally run generalist Vision-Language Models (VLMs) can detect FRBs in dynamic spectra under a zero-shot, prompt-only regime, with no fine-tuning and no labeled examples, returning structured decisions with a natural-language justification. From a controlled set of 3000 simulated L-band dynamic spectra containing FRBs, structured Radio Frequency Interference (RFI), and noise, we draw a balanced binary benchmark of 2000 samples and compare two such VLMs (Gemma 4 2B and 4B), sample by sample, against the state-of-the-art specialized detector SwinYNet. At the default threshold, Gemma 4 2B reaches an accuracy of 93.65%, with no statistically significant difference from SwinYNet (92.90%), while showing a significantly lower false-positive rate on structured RFI (6.4% vs. 25.0%) and no false positives on pure noise. SwinYNet retains a perfect probabilistic ranking on this benchmark (ROC-AUC of 1.0000 vs. 0.9482), a ceiling that the zero-shot VLM approaches from general-purpose pretraining alone. Rewriting the prompt alone reconfigures the same models for three-class FRB/RFI/noise classification on the full set of 3000 spectra, where they reach up to 86% accuracy without a single false FRB.
Raiff H. Santos, Amilcar R. Queiroz, Tharcisyo S. S. Duarte +2
Universidade Federal de Campina Grande, R. Apr´ıgio Veloso 882, Bairro Universit´ario, 58429-900, Campina Grande, PB, Brazil · Instituto de Forma¸c˜ao de Educadores, Universidade Federal do Cariri, Rua Oleg´ario Em´ıdio de Ara´ujo, s/n, Aldeota, 63260-000, Brejo Santo, CE, Brazil
The Nancy Grace Roman Space Telescope (Roman), set for launch as early as September 2026, will conduct wide-field infrared imaging surveys with unprecedented spatial resolution and cadence, enabling the discovery of millions of astronomical transients. Hence, it is necessary to have automated pipelines for generating alerts in place so that the telescope can begin discovering reliable transients and variable objects soon after it is launched. However, no real Roman data currently exist, making the development of such pipelines difficult. In this work, we present a machine learning model RuBR and a general methodology for distinguishing genuine transient and variable detections from spurious (bogus) detections within the RAPID pipeline. In particular, we present three models using this methodology: RuBRcomb trained and tested on combined locally injected and OpenUniverse2024 transients, RuBRloc trained on locally injected transients and tested on OpenUniverse2024 transients, and RuBRDA that combines locally injected transients with a fraction of OpenUniverse2024 transients in domain-adaptation mode for training. This paves the way for strategies to adapt the RuBRcomb model to real observations in the absence of any ground-truth labels during the early phases of the Roman mission. While the image differencing pipeline continues to be improved, our experimental results demonstrate the effectiveness of the proposed approach and its promise for robust real-bogus classification in the Roman era.
Karan Gandhi, Ashish A. Mahabal, Jacob E. Jencson +6
Department of Computer Science and Engineering, Indian Institute of Technology, Gandhinagar, India · Division of Physics, Mathematics, and Astronomy, California Institute of Technology, Pasadena, CA 91125, USA · Center for Data Driven Discovery, California Institute of Technology, Pasadena, CA 91125, USA +2
Radio frequency fingerprinting aims to identify transmitters by exploiting small differences in the emitted signal waveform. Common approaches exploit hardware-dependent imperfections for that purpose. This work instead investigates whether differences between digital signal processing implementations of a transmitter can be identified from their radiated signals. As a practical case study, five different software-based DSP implementations for generating a 2-FSK RTTY communication waveform are considered. In a monitoring receiver, a classifier can identify these DSP implementations using convolutional neural networks that are trained on augmented software-generated signals represented as raw IQ, power spectrum or instantaneous frequency. The trained networks are evaluated on over-the-air recordings obtained using a shortwave transceiver and a software-defined radio receiver. The best neural network combines IQ and the power spectrum representation and achieves an accuracy of up to 90%. The study further examines the influence of the signal SNR and the observation bandwidth on the classification accuracy. The results demonstrate that DSP implementations leave distinguishable characteristics in the transmitted signals and suggest that DSP-dependent fingerprinting can provide complementary information for specific emitter identification systems.