![HEP2026: 42nd Conference on Recent Developments in High Energy Physics and Cosmology [NTUA] Jul 8 – 11, 2026](/event/1654737/logo-1929213524.png){kind=logo}
Speaker
Description
PICOSEC Micromegas (PICOSEC-MM) is a high precision timing Micro Pattern Gaseous Detector (MPGD) developed to reach time resolutions on the order of a few tens of picoseconds. Its operation is based on detecting Cherenkov light generated in a crystal radiator. The ultraviolet photons produced are converted into photoelectrons by a photocathode, and the signal is subsequently formed through a two stage amplification process.
Test beam studies using muon beams have shown that PICOSEC-MM prototypes can achieve a time resolution of about $ \sigma_t \approx $ 10ps when the avalanches from all photoelectrons are collected on a single pad this is referred to as the intrinsic resolution. In addition, the signal formation mechanism in PICOSEC-MM detectors has been thoroughly investigated and modeled, providing a detailed understanding of their timing behavior, including resolution and systematic time offsets.
However, maintaining this level of precision becomes significantly more challenging in multi pad PICOSEC-MM detectors, where photoelectrons are distributed across multiple anode pads. Traditional signal reconstruction methods, based on combining individual pad signals, suffer from systematic errors. These effects are particularly pronounced for non central tracks, i.e., when the muon track crosses the detector near pad boundaries. Such systematic time offsets are difficult to model and typically require large calibration datasets for proper parameterization.
In this work, we present a novel signal reconstruction framework based on Graph Neural Networks (GNNs). We demonstrate that this approach can provide unbiased, high precision estimates of the particle arrival time using raw digitized waveforms as input. By representing the detector’s multi pad signals as a graph, the network learns both spatial correlations and signal sharing effects. Moreover, by employing a Gaussian Negative Log Likelihood (NLL) loss function, the model simultaneously performs regression of the particle arrival time ($t_{GNN}$) and estimation of its uncertainty ($\sigma_{GNN}$) on a track by track basis.
Our results, based on muon beam data, show that the GNN architecture effectively recovers the intrinsic timing resolution (within $\sim$ 1 ps), even in cases where photoelectrons are shared across pad boundaries. In addition, the model’s self estimated uncertainty is shown to be a consistent and unbiased measure of the true timing resolution. Finally, the computational efficiency of the GNN framework enables fast inference, offering a scalable solution for real time particle timing.