A hybrid neural-voxel physics-constrained framework for 3D reconstruction of radiation fields from multiple unknown sources

Accurate and rapid reconstruction of three-dimensional (3D) radiation fields is a critical task for nuclear emergency response and environmental monitoring. However, existing neural network reconstruction methods are heavily reliant on the quantity and quality of measurement data. Under sparse data conditions, this dependence often leads to physically implausible predictions, thereby compromising the reliability of safety assessments. To address this challenge, this paper introduces a Hybrid Neural-Voxel Physics-Constrained Framework (HNV-PCF). Unlike purely data-driven models, HNV-PCF embeds a discrete physical radiation transport model into a continuous neural network architecture, effectively coupling physical priors with data-driven rendering. Through comprehensive numerical simulations and high-fidelity scenarios, HNV-PCF was systematically compared with conventional data-driven methods under varying levels of data sparsity. Quantitative results demonstrate that, under typical sparse robotic sampling scenarios, HNV-PCF reduces the global Mean Absolute Percentage Error (MAPE) by 68.7% compared to the best-performing baseline. Furthermore, even in high-fidelity obstructed environments involving complex shielding, the proposed method robustly maintains physical consistency and accurately delineates the structural contours of hazardous regions. The method proposed in this paper offers an effective approach to solving the reconstruction problem of radiation fields under data-scarce conditions, demonstrating significant application potential in the field of nuclear safety.