An X-ray photon arrives with a position and an energy. Keep those together, and a model can ask whether emission of a particular energy is concentrated at a point or spread across the sky. Separate the observation into a picture and a spectrum, and part of that relationship is lost.[1]
A research team led by Maggie Lieu has tested these alternatives using simulated observations modelled on the European Space Agency's XMM-Newton X-ray observatory. Its paper, posted in October 2026, compared one-dimensional networks trained on spectral data, two-dimensional networks trained on images, combinations of the two, and a three-dimensional convolutional network that reads photon counts jointly in sky position and energy. The distinction is not just computational taste. It changes which relationships the machine can notice.[1][2]
A picture can lose a dimension
Imagine photographing a crowd in a station. A single image reveals where people stand, but not what each person said. An audio recording captures voices, but without positions it cannot always tell who spoke. An X-ray sky image and an energy spectrum have a similar information problem. The spatial distribution reveals whether emission is compact or spread out; the energy pattern distinguishes different physical mechanisms. Once the observation is reduced to independent summaries, some correlations may be lost.[1][2]
Galaxy clusters contain diffuse, extremely hot gas that emits thermal X-rays. An active galactic nucleus, powered by matter accreting on to a supermassive black hole, produces a different spatial and energy pattern. The benchmark tested whether models could distinguish simulated observations of those two source types.[1]
The investigation found that spectral resolution mattered, and that combining image and energy information was a strong step beyond treating either alone. The three-dimensional event-cube approach was the most consistent across the reported validation folds. But the authors did not claim a dramatic accuracy leap over the best combined two-dimensional and one-dimensional baseline. They described the gain there as marginal. The more interesting advantage was robustness and a reduction in preprocessing assumptions.[1]
The background that a human ordinarily removes
An X-ray detector also measures unwanted background counts. Conventional analysis includes careful estimates of that background and attempts to subtract it before modelling the source. Such work can require instrument expertise, particularly when the object is faint and the background uncertain. The three-dimensional network, working on total counts in the simulations, was able to learn useful distinctions without a separate, hand-crafted background model in its classification pipeline.[1][2]
Working from total simulated counts lets the network incorporate background characteristics within its classification. The transfer to real survey fields remains the key test, because the simulated sources and detector assumptions define what the network has learned.[1][2]
The researchers used three-dimensional Grad-CAM to inspect influential regions in the input. Its attention broadly followed physically recognisable differences between diffuse cluster gas and concentrated active nuclei, providing an interpretability check on the learned classification.[1]
Why simulated skies can still matter
The methodological gain is preserving the observation’s structure. A larger model is only one way to improve scientific AI; another is to avoid discarding relationships before the model gets to see them.[1][2]
Sources & further reading
- Data Representation Matters: Optimizing Machine Learning for X-ray Source Classification by Leveraging Spatio-Spectral Information arXiv · 8 Oct 2026 · Accessed 24 Oct 2026
- XMM-Newton mission European Space Agency · Accessed 24 Oct 2026