Towards DM-free search for Fast Radio Bursts with Machine Learning -- I. An implementation on multibeam data

TL;DR

This work introduces a DM-free FRB search framework that leverages an EfficientNet-based classifier operating on multibeam data to bypass exhaustive dedispersion. By training on simulated FRB/RFI data tailored to FAST's 19-beam layout and validating on real FAST observations, the approach achieves high discriminative performance (AUC > 97%, recall and precision around 92%) while naturally mitigating RFI. The method delivers substantial computational advantages, offering real-time-like throughput on GPUs and outperforming CPU-based pipelines in speed. A supplementary VAE-based DM estimation is proposed to recover dispersion information without dedispersion, and the authors discuss the viability of extending this DM-free paradigm to future all-sky, multi-beam surveys such as SKA.

Abstract

Searching for fleeting radio transients like fast radio bursts (FRBs) with wide-field radio telescopes has become a common challenge in data-intensive science. Conventional algorithms normally cost enormous time to seek candidates by finding the correct dispersion measures, of which the process is so-called dedispersion. Here we present a novel scheme to identify FRB signals from raw data without dedispersion using Machine Learning (ML). Under the data environment for multibeam receivers, we train the EfficientNet model and achieve both exceeding 92% accuracy and precision in FRB recognition. We find that the searching efficiency can be significantly enhanced without the procedure of dedispersion compared with conventional softwares like TransientX and presto. Specifically, the impact of radio frequency interference (RFI) for single-beam and multibeam data has been investigated, and we find ML can naturally mitigate RFI under the multibeam environment. Finally, we validate the trained model on actual data from the current FRB surveys carried out by the Five-hundred-meter Aperture Spherical radio Telescope, which provides considerable potential for real implementation in the future.

Figures (12)

• Figure 1: The layout of the FAST 19-beam receiver Jiang+20RAA. The pointing system is centered on Beam 01, with Beams 02-19 surrounding the central beam.
• Figure 2: The architecture of the model used in this work. The input data is preprocessed to 224 pixels $\times$ 224 pixels $\times$ 7 channels, the EfficientNet model Tan+19arXiv is used as the base model, the output feature of the fully connected layer of the model is modified to 256, and an additional Rectified Linear Unit (ReLU) activation function layer, Dropout layer, and a fully connected layer with an input feature number of 256 and an output feature number of 2 are added. Finally, the predicted output of the image classification is obtained.
• Figure 3: Training samples of multibeam datasets. From left to right, the types of RFI are impulsive, point-to-point, satellite, and narrowband. In each subfigure, RFI is distributed across 7 beams, while FRB only appears randomly in one beam. In the panels of each sub-figure, the x-axis and y-axis represent time and frequency, respectively. Note that the samples in the multibeam dataset are formed by combining the central 7 beams into 7 channels used in Figure \ref{['fig:model']}.
• Figure 4: Confusion matrices of the multibeam model. From left to right, the first row is TN, FN, and the second row is FP, TP. N or P denotes whether the model predicts the sample to be a positive sample or a negative sample. T or F indicates whether the model is correct or wrong.
• Figure 5: Multibeam ROC under different RFI cases. The $ROC\_overall$ plot shows AUC values $>$97%. The model achieves the AUC exceeding 98.5% under both satellite and Point-to-Point RFI conditions. The AUC also exceeds 94% under the other two RFI conditions, indicating that the multibeam model is suitable for FRB searches in the presence of strong RFI.