Achieving precise identification of neutrons/gammas in space is beneficial for studying space radiation fields. Traditional identification methods are significantly affected by data noise, making it challenging to distinguish neutrons/gammas under the high-noise conditions of spaceborne detectors.

This study aims to propose a deep learning algorithm that combines Scattering Convolution Networks with MobileNetV4 (MNv4) and Kolmogorov–Arnold Networks (KAN) to improve neutron/gamma discrimination performance and the noise robustness of this method.

Firstly, a ?50 mm×50 mm NaI(Tl):⁶Li crystals and H6410 photomultiplier tubes were employed to collect data on neutron beams of 2 500 keV and 5 000 keV from the National Defense Technology Industrial Ionizing Radiation Calibration Station at China Institute of Atomic Energy (CIAE). Then, the 2 500 keV data were divided into test set 2 and training set, while the 5 000 keV data served as test set 1. After classifying the dataset using the charge comparison method, the scattering convolution networks (SCN) was used for denoising and dimensionality reduction, reducing 1 800 sample points to 113. Particle identification was then performed using MNv4 or KAN model. Finally, simulated signals with noise levels of 10%~50% were applied to investigating the algorithm's noise resistance and tested different signal-to-noise ratios with falling edge trigger leves of -4 mV, -8 mV, and -12 mV to verify the algorithm's robustness.

The experiment results show that MNv4 and KAN achieve recognition accuracies of 99.7% and 98.6% on the two test sets, respectively, compared to the comparison method, which improves the Figure of Merit (FOM) from 1.82 to 4.07 and 4.40. At a 15% noise level, MNv4 and KAN achieve neutron identification accuracies of 99.7% for 2 500 keV and 5 000 keV energies whilst the identification accuracies of above two algorithms are maintained above 99% with a signal-to-noise ratio (SNR) of 9.13 dB at a 40% noise level.

Combining SCN with deep learning algorithms like MNv4 or KAN provides dimensionality reduction and noise resistance. Compared to the charge comparison method, this approach improves the Figure of Merit (FOM) and enables neutron/gamma discrimination even at high noise levels, demonstrating good robustness.