Noise reduction for self-emission x-ray images in ICF using a physics-informed deep learning framework

Self-emission x-ray imaging is a key diagnostic in inertial confinement fusion (ICF), yet the recorded images are heavily degraded by shot noise arising from photon-counting statistics and scintillator blur, as well as impulsive noise induced by neutron interactions with silicon-based detectors. The complexity of these noise sources, together with the limited availability of experimental data, poses significant challenges for both physical interpretation and data-driven denoising. To address this problem, we propose a hybrid physics-informed and generative framework that enables realistic data synthesis and robust denoising of self-emission x-ray images from ICF experiments. Hot-spot signals are modeled using Legendre polynomial expansions, while shot and impulsive noise are synthesized from empirically measured distributions. A generative enhancement module further reduces the domain gap between synthetic and experimental data, resulting in more realistic training inputs and improved denoising performance. Self-emission x-ray images from experiments conducted on SG-180 kJ laser facility demonstrate the effectiveness of the proposed approach. The synthesized data achieve peak signal-to-noise ratio values reaching 61 dB, and the denoising results improve the signal-to-noise ratio of real experimental images from approximately −30 dB to +35 dB.