Crosstalk mitigation in an ATLAS-like high energy liquid argon calorimeter using artificial neural networks

High-energy particle accelerators and their associated experiments contribute significantly to our understanding of the nature of matter. In this context, calorimeters are key components employed to measure the energy of the outgoing particles produced in the collisions. For this, a high-density structure of sensors organized around the beam axis and comprising several superimposed detection layers is used. Calorimeter information is essential for particle characterization and triggering. A high-energy sampling calorimeter is split into electromagnetic and hadronic sections and constructed using an alternate structure of sensitive and absorber materials. One possible configuration for electromagnetic calorimeters is to use liquid Argon (LAr) and lead (Pb) for sensitive and absorber materials, respectively. The high-density granular structure of sensor cells, combined with high energy values of the particles to collide, can produce parasitic effects that distort the signal of interest in each cell. One of these effects is crosstalk (XT), a spurious signal induced from neighboring sensor cells. Consequently, particle characterization is affected, introducing errors in energy and particle time-of-flight estimation. This work proposes a machine-learning method for crosstalk mitigation in an ATLAS-like LAr+Pb electromagnetic calorimeter. For this, a study based on a framework to simulate particle interaction on an LAr calorimeter that accounts for the crosstalk effect as observed experimentally was proposed. The results show that a supervised learning neural network model for time regression reduces the estimation error by 31 times for cells with the highest energy and 182 times for cells with low energy. For energy estimation, the neural estimator reduces the error by 1.09 times at high-energy cells and 12 times at low-energy cells. The results also indicate that the proposed model is capable of reducing the crosstalk distortion in shower-shape variables, which are traditionally used for particle identification in calorimetry. The implementation of such a method at LHC experiments could provide a very interesting improvement to identify particles not associated with a collision of interest (pileup) based on an enhanced time of flight measurement. • A simulation modeling for crosstalk in an ATLAS-like calorimeter is proposed. • Crosstalk signal distortion is significantly reduced with a machine learning solution. • The proposed method reduces the time-of-flight estimation error by up to 182 times. • Energy estimation error is reduced by up to 12 times for low-energy cells. • The method significantly reduces the shower-shapes distortion caused by crosstalk.