Neural networks read a fast radio burst's dispersion straight from its spectrum

The work

Technical · from the record

Three deep-learning models — a baseline CNN, a fine-tuned ResNet-50 and a hybrid CNN–LSTM — were trained to read the dispersion measure of a fast radio burst directly from its frequency–time dynamic spectrum. Training and validation used 180,000 synthetic CHIME/FRB-like spectra in six-fold cross-validation, spanning dispersion measures of 0 to 600 pc cm-3, with a separate test set of 10,000 samples. On the fold-averaged test set the hybrid CNN–LSTM gave a mean absolute error of 0.2542 pc cm-3, ResNet-50 0.4802 and the baseline CNN 0.6690, and 98.06 per cent of hybrid-model predictions fell within the 1.1 pc cm-3 absolute-error threshold the authors adopted. Preliminary application of the trained models to a sample of high signal-to-noise CHIME/FRB bursts recovered dispersion measures but did not reach that threshold as often as on simulated data.

How AI was used

Synthetic frequency–time waterfalls were simulated with CHIME/FRB-like instrumental settings (400–800 MHz, 600 MHz reference frequency, 1.67 ms effective sampling, 512 frequency channels) using an extended version of a public simulation package, adding dispersion delay, a TRDM sub-burst drift term, exponential scattering, scintillation, a Euclidean fluence distribution, a uniformly sampled spectral index and Gaussian noise. The resulting labelled dataset was split into six cross-validation folds plus an independent test set, and preprocessed per model: spectra resized to 224×224 for the baseline CNN, and for ResNet-50 additionally given extra uniform background noise, min–max scaled, replicated to three channels and standardised with ImageNet statistics. Three architectures were then fitted as scalar regressors on the true dispersion measure: a seven-layer convolutional network with five dense layers trained from scratch with an MAE loss and Adam; ResNet-50 with its ImageNet classification head replaced by a single linear output, its early layers frozen and its final two convolutional stages unfrozen, trained with AdamW, MAE loss and a step learning-rate scheduler; and a hybrid network applying one-dimensional convolutions along the time axis followed by two stacked bidirectional LSTM layers and a dense regression head, trained with an MSE loss. Dropout was omitted in both the baseline CNN and the hybrid model. Trained models were run over the validation folds and the held-out synthetic test set, with per-sample predictions averaged across folds after dropping the extreme values, and were also run on a sample of high signal-to-noise CHIME/FRB bursts. Non-learned steps computed regression metrics, threshold-based accuracy rates, kernel density estimates of the error distribution, and training and inference timings.