Optimization of Energy Resolution in Topmetal-II‾ Pixel Detector: An Exploration of Neural Network and Curve-Fitting Strategies

This study aims to enhance the energy resolution of the Topmetal-II $^{-}$ pixel detector, which is pivotal for capturing soft X-ray signals and accurately measuring X-ray polarization. To achieve this, we used convolutional neural networks (CNNs) and curve-fitting techniques. The CNNs were trained on energy data from the Topmetal-II $^{-}$ detector to predict the peak voltage of waveforms at different amplitudes. In addition, we employed curve-fitting methods to develop a response model for the detector. We collected energy data from the Topmetal-II $^{-}$ using high-precision oscilloscopes, with data sampled at equidistant intervals to simulate a 40-MHz sampling rate analog-to-digital converter (ADC). Our findings indicate that a denoising autoencoder (DAE) with five encoding layers and five decoding layers, combined with a two-layer regression network, significantly improved performance. The energy deviation was reduced from 15.4 to 0.41 mV, and the energy resolution was enhanced from 24.26 to 5.09 mV, representing an approximate 4.77-fold improvement. In contrast, the curve-fitting increases the energy peak from 27.04 to 32.05 mV at an input of 350 mV, resulting in a slightly improved energy resolution. These results demonstrate substantial advancements in soft X-ray detection technology, underscoring the potential for more precise and reliable measurements in low-energy X-ray polarization detector (LPD) experiments.