Machine-Learning-Assisted Optimization of Pulse Shape Discrimination in Gadolinium-Loaded Liquid Scintillators for the RENE Experiment

Pulse shape discrimination (PSD) is crucial for distinguishing neutron- and gamma-induced events in liquid scintillator detectors. We assess PSD performance using gadolinium-loaded liquid scintillator (Gd-LS) with 10% by volume di-isopropylnaphthalene (DIN) additive, exposed to a 252Cf source in a controlled setup. Our study compares traditional methods-charge ratio (Qratio) and log-likelihood ratio (LLR)-with machine learning approaches, including principal component analysis (PCA) and an autoencoder (AE) combined with PCA. An unsupervised AE compresses 150-dimensional waveforms into a 12-dimensional latent space, further reduced to 2D via PCA. Pure neutron and gamma samples, selected using extreme Qratio values, serve as training data for supervised classifiers. The AE+PCA method achieves over 99% classification accuracy, with a figure-of-merit (FoM) of approximately 3.20 and a fit fidelity (FF) of about 0.97, outperforming conventional techniques. These results demonstrate the effectiveness of nonlinear feature extraction and dimensionality reduction in PSD tasks. The proposed approach offers a robust framework for enhancing event classification in future neutrino experiments like RENE, where precise neutron background suppression is essential.