Speaker
Description
Resistive Plate Chamber (RPC) detectors are widely used in major CERN experiments as muon trigger systems thanks to their excellent time resolution. However, the gas mixtures currently employed in RPCs contain greenhouse gases, such as C2H2F4 (R134a) and SF6, with Global Warming Potential (GWP) of 1430 and 22800 respectively, which contribute significantly to environmental emissions. During LHC Run 2, approximately 85% of the emissions from particle detectors originated from RPC gas leaks. For this reason, several environmentally friendly alternative gas mixtures have been investigated. Building on these efforts, this work aims to develop a neural-network-based model capable of predicting RPC efficiency curves for low-impact gas mixtures. A feed-forward neural network with two hidden layers is trained to predict the three parameters of the sigmoid efficiency curve, maximum efficiency, slope, and HV at 50% efficiency, as a function of the gas mixture composition. The model is trained on experimental data collected at the CERN Gamma Irradiation Facility (GIF++) using 16 alternative gas mixtures based on HFO and CO2 as R134a substitutes, and Novec 4710 as an SF6 alternative. Model performance is evaluated using Leave-One-Out cross-validation. Preliminary results show good agreement between model predictions and experimental data when the gas composition lies within the parameter space covered by the training dataset. The goal is to extend the model's predictive power to mixtures with arbitrary combinations of known components, as well as new gas components not included in the training data, thus providing a versatile tool for the study and optimization of sustainable RPC operation. This tool is particularly relevant in view of the upcoming Long Shutdown 3, during which experimental measurements at GIF++ will not be possible, making predictive modelling a key asset for advancing the search for sustainable gas mixtures. This poster will present the current status of the project.