A 5 m Cherenkov telescope with transformer analysis drops its energy threshold to 0.18 TeV using LST-1 photomultipliers, against 0.73 TeV for the standard analysis
Synopsis
In simulations of an idealised 5 m stand-alone imaging atmospheric Cherenkov telescope, a video vision transformer reads the raw camera waveforms instead of Hillas parameters computed after image cleaning. With the photomultipliers actually used in LST-1, the standard analysis raises the energy threshold to 0.73 TeV while the transformer analysis reaches 0.18 TeV; the standard chain would need 100% quantum efficiency to match that threshold. The reason is that the transformer works with events of about 35 photoelectrons in the camera, whereas the standard analysis needs about 200 before cleaning.
Interpretation
On the same 5 m telescope and the same simulated events, changing only the analysis separates the two chains at low energy: the transformer threshold rises far more slowly as photons are lost. Event reconstruction for such telescopes has mostly followed the standard analysis: each pixel waveform is reduced to one charge and one arrival time, pixels likely containing only night-sky light are removed, and a few Hillas image parameters feed the reconstruction, nowadays often random forests. Going from 100% quantum efficiency to the LST-1 photomultipliers, the transformer threshold rises from 0.074 to 0.18 TeV and the standard threshold from 0.19 to 0.73 TeV; at 0.5 TeV the standard sensitivity degrades by a factor of 19 and the transformer by 2.2.
The transformer analysis takes the whole waveform movie as input, and one network returns the energy, two arrival-direction coordinates in the camera plane and a gamma/hadron score. The network is a factorised-encoder video vision transformer, cutting the movie into tubelets of several pixels and two time samples and turning each into a token. A spatial encoder merges the tokens of each pair of time samples, a temporal encoder processes the resulting sequence, and each configuration is trained separately with the earlier hyperparameters, taking about 7.5 hours on four NVIDIA A100 GPUs.
The low-energy gain comes from how much signal each chain uses: events in the transformer's lowest energy bin carry about 35 photoelectrons in the camera on average, while the standard analysis's lowest bin carries about 200, of which about 80 remain in the image after cleaning. The standard analysis integrates charge in a short window around the pulse peak, and image cleaning keeps only pixels above a threshold plus qualifying neighbours, so the image size is 40–70% of the camera photoelectrons. With the LST-1 photomultipliers, the transformer analysis reaches the 0.063–0.1 TeV bin while the standard analysis has no usable bin below 0.25 TeV; their thresholds are 0.18 and 0.73 TeV.
Perspective
The conclusions rest on an idealised stand-alone telescope generated with the CTAO simulation tools: a perfectly reflecting and aligned mirror, no electronic noise, a dark-sky background, and a proton-only background simulated only within the pointing-direction cone and 250 m of the telescope. In the comparison with MAGIC and LST-1, those sensitivities come from real data, so the curves are not fully comparable; the authors frame the comparison as context for the simulated result.
The authors state that the simulated background is probably a few tens of per cent too low, but because the weakest detectable flux scales only as the square root of the background where background dominates, the effect on sensitivity is smaller; each idealisation favours the simulated telescope, so absolute sensitivities are optimistic. Network hyperparameters were not optimised for any configuration and the gamma/hadron cut is fixed, while retraining with different random seeds changes sensitivity by at most about 12%, so differences below that size between configurations are not interpreted.
