Neural networks trained on simulated radio bursts sort real bursts by shape

The work

Technical · from the record

The authors built a simulation framework around the simpulse library to generate dynamic spectra for six fast radio burst morphology archetypes, generating 1000 samples per type at each of five signal-to-noise values, and used these to train convolutional neural network classifiers. Two schemes were trained: ten pairwise binary classifiers whose threshold-subtracted confidences are summed into a five-class decision, and a single multi-class network trained on simulated bursts together with half of the first CHIME/FRB catalog. Tested on real catalog bursts, the single multi-class classifier reached an overall accuracy of approximately 55%, against a random multiclass rate of 20% for five balanced classes, while its accuracy on the simulated test set was about 73% and the binary-model scheme reached roughly 50% on catalog data. SHAP attributions were computed over catalog bursts to inspect which regions of the dynamic spectra contributed to predictions.

How AI was used

Because published FRB samples are few and unevenly distributed across morphologies, the training data were synthesised: a wrapper around the simpulse library composed bursts with specified numbers of components, spectral shapes (power law, Gaussian, diffraction pattern), scattering, dispersion-measure error and Gaussian noise, sampling parameters from uniform ranges plus a log-normal width distribution taken from the first CHIME/FRB catalog, over a 400-800 MHz band at 1 ms sampling and rendered as 256x256 arrays. Bursts whose components overlapped indistinguishably were removed by hand, and the remainder split into training, validation and test sets drawn from a mixture of signal-to-noise levels. Ten CNN binary classifiers, one per pair of five classes, were trained from scratch in keras and tensorflow with ReLU activations, dropout, Adam and early stopping on validation accuracy; kerastuner RandomSearch over dense-layer units, learning rate, batch size and dropout selected final models, and a per-classifier decision threshold was chosen where false positive and false negative rates meet. At inference each spectrum passes through all ten classifiers and the threshold-subtracted confidences are summed per row to give a consensus class. Separately, a larger single network with a five-way softmax was trained on simulated bursts plus half of the CHIME/FRB catalog, with the other half held for testing. Real catalog waterfalls were rebinned from 16384 to 1024 channels, masked channels filled by resampling neighbouring valid intensities, and padded with matched Gaussian noise before classification. SHAP was applied to the single classifier to attribute predictions to regions of the input.