Neural network-based nuclear charge Z identification from ionization chamber waveforms of low-energy heavy ions

This study investigates Z identification by analyzing the low-energy heavy ion waveforms in a high-resolution low-pressure ionization chamber through a multilayer perceptron (MLP) neural network. Five nuclides - 127I, 128,129,134Xe, and 133Cs - were measured and their waveforms normalized to minimize interference from mass and energy. The MLP model, trained on these normalized waveforms, achieved a skill score of 53.2%. Our analysis of the amplitude distribution of waveform sampling points revealed that the MLP identified Z by considering the amplitude of the sampling points where statistical characteristics are learned during training. Furthermore, the skill score increased with the number of statistical characteristics learned. Therefore, three methods were developed to enhance the skill score: mean waveform training dataset generation, training dataset mean waveform augmentation, and training dataset waveform downsampling, which increased the skill score to 75.2%, 70.5%, and 71.5%, respectively.