Repeated views of the sky help AI recognise a telescope’s distortion

A 27 October preprint tests physical repetition in simulated observations: a fixed telescope watches different sky positions during a scan, then sees the same sky under changed instrument conditions. Pairing those views trains the network to represent sky and telescope separately, without labelling instrument settings.[1]

The reason this matters is visible in an earlier CHIME analysis of real observations. Bright foreground radio emission overwhelms the cosmological hydrogen signal. Its smooth variation with frequency helps astronomers separate the two, but instruments and processing can introduce false spectral structure.[2]

The CHIME team’s November 2025 preprint illustrates the trap with two individually smooth daily scans. Give one 1% greater amplification and remove different frequency channels from each: averaging the surviving measurements creates an uneven spectrum. Their pipeline therefore filters foregrounds before averaging observations, and carries the filter into simulated hydrogen signals to track what the processing removes. Repetition helps only when the changing measurement process is accounted for.[2]

The new AI test varied one ripple in the beam’s width across radio frequencies. After training on 70 instrument states, a labelled linear calibration predicted its amplitude across 15 unseen states with R² of 0.9975, close to a perfect fit of 1, averaging 78 held-out sky positions.[1]

However, exchanging sky representations between different instrument states increased median reconstruction error about 2.5-fold, exposing residual telescope contamination. Nearby test pointings were correlated with training positions; performance weakened on an unseen sky region. Thermal noise remains untested. Identifying one instrument parameter has not yet recovered the faint cosmological hydrogen signal.[1]

Report an error or suggest a correction