Machine-learning approaches to dispersion measure estimation for fast radio bursts

TL;DR

This paper addresses the challenge of accurately estimating the dispersion measure (DM) of fast radio bursts (FRBs) in a data-driven manner suitable for large surveys. It compares three deep-learning architectures—a baseline CNN, a fine-tuned ResNet-50, and a CNN–LSTM hybrid—trained and validated on a large synthetic CHIME/FRB-like dataset to recover DM directly from frequency–time dynamic spectra. The CNN–LSTM model delivers the best accuracy and robustness (MAE ≈ 0.25 pc cm$^{-3}$, RMSE ≈ 0.64 pc cm$^{-3}$) with favorable computational efficiency, suggesting real-time applicability and scalability for FRB surveys; ResNet-50 provides a strong intermediate performance, while the baseline CNN trails behind. The results support a data-driven DM estimation pathway that complements traditional dedispersion methods and can be adapted to real CHIME/FRB data or next-generation facilities such as the SKA.

Abstract

Fast radio bursts (FRBs) are bright, mostly millisecond-duration transients of extragalactic origin whose emission mechanisms remain unknown. As FRB signals propagate through ionized media, they experience frequency-dependent delays quantified by the dispersion measure (DM), a key parameter for inferring source distances and local plasma conditions. Accurate DM estimation is therefore essential for characterizing FRB sources and testing physical models, yet current dedispersion methods can be computationally intensive and prone to human bias. In this proof-of-concept study, we develop and benchmark three deep-learning architectures, a conventional convolutional neural network (CNN), a fine-tuned ResNet-50, and a hybrid CNN-LSTM model, for automated DM estimation. All models are trained and validated on a large set of synthetic FRB dynamic spectra generated using CHIME/FRB-like specifications. The hybrid CNN-LSTM achieves the highest accuracy and stability while maintaining low computational cost across the investigated DM range. Although trained on simulated data, these models can be fine-tuned on real CHIME/FRB observations and extended to future facilities, offering a scalable pathway toward real-time, data-driven DM estimation in large FRB surveys.

Figures (6)

• Figure 1: Example of a simulated FRB dynamic spectrum with $\mathrm{DM}\approx559.388~\mathrm{pc\,cm^{-3}}$.
• Figure 2: Activation flow through the hybrid CNN–LSTM model during DM estimation. The model input is a waterfall plot of size $512{\times}1024$, representing frequency (512 channels) and time (1024 samples). The CNN front end consists of four one-dimensional convolutional layers that apply multiple filters comoving along the time axis for each frequency channel, producing a compact sequence of feature maps of size $(256,128)$. This feature sequence is passed to the recurrent back end, composed of two bidirectional LSTM (BLSTM) layers. In each BLSTM layer, the feature sequence is processed in both forward and backward temporal directions, generating hidden states $H_\mathrm{F}$ and $H_\mathrm{B}$ with 128 units each. The hidden states from both directions are concatenated along the feature dimension at each time step and passed to the next BLSTM layer, enabling the network to capture dependencies extending in both temporal directions. The fully connected regression head receives input from the final concatenated hidden states of the second BLSTM layer and produces a single scalar output corresponding to the estimated DM. This structure enables the model to integrate local spectro–temporal features extracted by the CNN with long-range temporal context learned by the bidirectional LSTM layers.
• Figure 3: Learning curves for the best-performing fold of each model. The panels show the mean absolute error (MAE) of the training and validation datasets as a function of epoch number. The baseline CNN was trained for 60 epochs, whereas the ResNet-50 and hybrid CNN–LSTM models were trained for 150 epochs, taking advantage of shorter epoch durations and smaller memory footprints.
• Figure 4: Each panel shows the histogram of absolute errors $|\Delta \mathrm{DM}|$ (in pc cm$^{-3}$) for 10,000 test samples; the vertical axis gives the count of predictions per bin on a logarithmic scale. The baseline CNN exhibits the broadest spread (66 outliers with $|\Delta \mathrm{DM}|>2.8$ pc cm$^{-3}$; maximum $\sim$42 pc cm$^{-3}$), ResNet-50 is narrower (five outliers; maximum $\sim$4.5 pc cm$^{-3}$), and the hybrid CNN--LSTM shows the highest concentration near small errors but a slightly larger number of outliers and a higher maximum error than ResNet-50 (15 outliers; maximum $\sim$15.7 pc cm$^{-3}$).
• Figure 5: Absolute error as a function of the true DM for the three models. Each scatter plot shows 10,000 test samples, with points color-coded by their kernel density estimate (KDE) to highlight regions of high prediction density. Two solid lines in green and cyan denote the adopted error thresholds: an absolute error of 1.4 pc cm$^{-3}$ and a relative error of 1%, respectively. The hybrid CNN--LSTM exhibits the most symmetric and tightly clustered distribution across the full DM range, demonstrating uniform accuracy and minimal systematic bias.