Advancing background rejection capabilities in the CYGNO experiment via machine learning

Abstract

The direct detection of Dark Matter requires the identification of rare Nuclear Recoil (NR) interactions in the presence of dominant Electron Recoil (ER) backgrounds. The CYGNO experiment addresses this challenge through a gaseous Time Projection Chamber with optical readout, capable of reconstructing the topology of ionization tracks produced by particle interactions. This work presents the development and evaluation of machine learning frameworks for ER/NR discrimination using simulated data from the LIME prototype. A complete analysis pipeline was established, including image reconstruction, feature extraction, detector response characterization, Monte Carlo simulation tuning, and classifier optimization. A set of 30 topological and energy-related observables was extracted from reconstructed tracks and used to train classical machine learning algorithms, while image-based deep learning models were trained directly on the detector images. The optimized classifiers were evaluated using cross-validation and an independent test set, achieving a maximum ROC AUC of 0.902. A detailed energy-dependent analysis showed that discrimination performance is strongly limited at low energies and approaches perfect separation at higher energies. In the intermediate energy region, where the classification task is most challenging, feature-based approaches achieved background rejection factors comparable to or exceeding those of the image-based models. Among the evaluated methods, XGBoost and a Multilayer Perceptron provided the best combination of discrimination performance and computational efficiency, reaching inference rates above 105 events per second while maintaining competitive rejection power. The results demonstrate that the engineered observables successfully capture the relevant physical information contained in the ionization tracks, allowing classical machine learning models to match and, in some operating regimes, exceed the performance of convolutional neural networks. Overall, the study indicates that feature-based machine learning provides an efficient solution for ER/NR discrimination in the LIME detector while preserving the possibility of interpreting the physical observables responsible for the classification.