A DM-free scheme for fast radio burst search in multibeam data based on EfficientNet

Chen, Yao

Fast Radio Bursts (FRBs) are high-energy astrophysical phenomena of cosmological origin， characterized by transient radio pulses lasting only a few milliseconds. Currently， the identification of FRBs signals with traditional method primarily relies on dedispersion manual image-by-image recognition， which requires significant human and time resources. In the recent years， with rapid development of deep learning on computer vision， it becomes particularly import to develop more efficient ways on FRB blind search based on Machine Learning.

Here we present a new scheme on searching for FRBs in the multibeam data without dedispersion based on EfficientNet， which is a novel convolutional neural network architecture increasingly applied in deep learning. Using the scaling method， the Efficient-Net model would enhance network performance by simultaneously increasing the width and depth of the network and the resolution of input images. To demonstrate this new scheme for FRB search， we first simulate 1-bit PSRFITS data using the software “simulateSearch” in the observational environment similar with the FAST 19 beams. Then we train the EfficientNet model with the 19-beam labelled data simultaneously， in a GPU cluster composed of RTX 4090 graphics cards. After that， we benchmark the performance of this AI-based method by running the pipelines based on commonly used single-pulse search softwares， such as Heimdall， TransientX and PRESTO. In summary， this new scheme can enhance the efficiency for FRB blind search compared with other traditional algorithms. Additionally， this approach would naturally mitigate the negative impact by Radio Frequency Interference (RFI) in multibeam data.