This paper presents a design for a high-speed digital prototype system for discriminating fast neutrons, thermal neutrons, and γ-rays. The system uses a stilbene–⁶Li glass composite scintillator with excellent pulse shape discrimination (PSD) properties as the neutron detector. The PSD performance was investigated at different sampling rates, revealing stable performance at rates above 250 MSPS. The system core is a high-speed acquisition board based on the AD9434 analog-to-digital converter (ADC) and the ZYNQ7020 field-programmable gate array (FPGA), which acquires detector signals and implements real-time algorithms. The system was energy-calibrated with ²²Na, ¹³⁷Cs, and ⁶⁰Co γ-ray sources and evaluated in a n–γ mixed field. Under an ²⁴¹Am–Be neutron source, the system achieved Figure of Merit (FOM) values of 1.26 for fast neutron/γ, 2.18 for fast neutron/thermal neutron, and 3.25 for γ/thermal neutron discrimination above the 50 keVee electron equivalent energy threshold. These results are consistent with the analysis of down-sampled data from a DT-5730 digitizer, confirming that the system meets its design objectives. Additionally, the measured false alarm rates (FAR) were 0.33% for ⁶⁰Co, 0.34% for ¹³⁷Cs, and 0.26% for ²²Na. This system integrates waveform discrimination and energy spectrum measurement capabilities, providing a high-performance, cost-effective electronic solution for high-speed signal acquisition and real-time processing in novel composite scintillator neutron detectors.

Predicting the luminescence behavior of Cs2ZrCl6 crystalline powders with speed and accuracy is a key prerequisite for designing new scintillators aimed at X-ray imaging. Conventional first-principles calculations based on density functional theory can reproduce crystal luminescence, yet they demand substantial computing power and deep materials-domain knowledge, which together translate into high cost, long development cycles, and limited throughput. To sidestep these bottlenecks, our work adopts a machine learning strategy to predict the luminescence properties of Cs2ZrCl6. A training dataset was assembled from the literature and expanded through linear interpolation. The hyperparameters of six regression algorithms, including ridge regression (RR), K-nearest neighbor regression (KNR), support vector regression (SVR), random forest regression (RFR), XGBoost regression (XGBR), and multilayer perceptron regression (MLPR), were tuned by coupling the covariance matrix adaptation evolution strategy (CMA-ES) with 10-fold cross-validation. Once trained, the models captured a reliable composition-structure-property link. By comprehensively evaluating three metrics of the coefficient of determination (R²), mean absolute error (MAE), and mean squared error (MSE), two optimized luminescence property prediction models were identified. The KNR model was selected for emission wavelength prediction, achieving a high R2 of 0.9999 and a low MSE of 1.4434. Meanwhile, the MLPR model was chosen for decay time prediction, with a high R2 of 0.9893 and a low MSE of 0.9234. Additionally, the Shapley Additive exPlanations (SHAP) analysis further quantified feature contributions to target properties, offering physically reasonable interpretations. This work demonstrates the promising capacity of machine learning for fast, reliable prediction of crystal luminescence performance, providing practical support for accelerating materials-oriented research and development.