A neural network replaces the FFT-based image reconstruction used by the Swift BAT telescope and localises X-ray sources with over 90 per cent accuracy

Tested in a simulation using Swift’s Burst Alert Telescope, an AI system directly deduces the positions of hard X-ray sources from detector counts, without knowing the pattern of the code-sliced image.

The X-ray telescopes coded mask do not use mirrors: a perforated mask casts a shadow on the detector, and an image must then be reconstructed by calculation, traditionally using fast Fourier transforms. Two researchers propose to entrust this stage to a network of neurons capable, in time, of turning in real time on board a spacecraft.

Their demonstration, conducted on simulated responses of the Burst Alert Telescope (BAT) Swift satellite, combines extraction of characteristics by convolution and attention mechanism. Driven around 900 photons per source, the model locates the source at the exact pixel in more than 93% of cases with only 400 photons detected in the absence of background noise, and remains above 90% for a 900 photons source drowned in an additional 8,500 hits.

Learn to "see" without knowing the mask

Note: the network does not need to be explicitly provided with the pattern of the mask, it learns from the data. The authors see it as a path to instruments with non-plane geometries, impossible to exploit with conventional methods, and towards better performance for telescopes already in operation.