deepclean3d
De-noising the Large Hadron Collider.
Enabling the search for physics beyond the standard model.
De-noising the Large Hadron Collider.
Enabling the search for physics beyond the standard model.
TORCH is a sub-module of the Large Hadron Collider, designed to provide particle identification below 10GeV/c and scheduled for installation at LHCb during the LS3 upgrade.
TORCH is made up of a bank of photon-multiplier tubes arranged in a grid, similarly to the pixels of a cameras CCD. While a typical camera operates in the range of 30-120fps, TORCH's PMTs readout at a rate of twenty billion fps, yielding a temporal resolution of 50 picoseconds per frame.
You may be asking why are we sensing photons when the goal is to identify hadrons. TORCH contains a thin sheet of quartz that is placed in the hadrons path, while the hadrons are transiting through the quartz they emit photons through the process known as Cherenkov Radiation.
The photons travel through the quartz by means of total internal reflection until they arrive at the PMT arrays, which are attached along the edges of the quartz. Analysis of the digitised output signal is used to reveal the momentum of the hadron, from which the identity is inferred.
efficiency
To perform the analysis, the path that each photon took through the detector must be fully reconstructed, a computationally costly process. The compute time of the reconstruction scales linearly with the number of photons detected. Of the photons detected, over 99% originate from overlapping tracks or sources of noise and are discarded post reconstruction, leading to a low efficiency.
DEEPCLEAN3D uses a deep neural network to remove uncorrelated photons from the volumetric PMT array data pre-reconstruction, greatly enhancing the efficiency.
We focus on three direct performance metrics: decreasing the uncorrelated photons in the data, retention of correlated photons and minimal introduction of false positives through processing artefacts that can cause erroneous reconstruction.
Noise Reduction
Signal Retention
False Positives
Efficiency
Our latest model is available for testing below, the current version is still in early alpha and not yet ready for deployment. For help or information on how to prepare your dataset and finetuning your training please consult the documentation.
V0.94documentation
