Geometry-aware and physics-informed 3D radiation field reconstruction from sparse observations

Accurate perception of three-dimensional radiation environments is essential for nuclear accident response and nuclear facility decommissioning. This paper presents a mobile robot-based method for 3D radiation field reconstruction. Unlike conventional interpolation approaches, the proposed approach establishes a physics-guided framework in which source localization supports radiation field mapping. An autonomous patrol-to-source-seeking switching strategy is first designed using a mechanical collimator. Angle-constrained particle filtering and maximum likelihood estimation are then employed to rapidly identify source parameters. These estimates provide accurate physical priors for subsequent field reconstruction. A physics-informed heteroscedastic Gaussian process model, termed PI-HGPR, is further introduced for radiation field inference. The model captures fluctuations in high-radiation regions by incorporating heteroscedastic noise. It also incorporates a physics-based attenuation kernel to reduce oversmoothing under sparse sampling. Finally, the reconstructed continuous radiation field is fused with a 3D geometric map to generate a radiation point cloud. Experimental results demonstrate that the proposed method maintains high reconstruction accuracy under sparse-sample conditions, achieving an RMSE below 7%, and provides a reliable data foundation for unmanned nuclear emergency exploration.