Direction and distance affect both the magnitude and spatial distribution of counts in a compact scintillator array, which makes their simultaneous estimation difficult. We modeled radiation transport and detector response in Geant4 11.3.2 for a 12 × 12 yttrium orthosilicate (YSO) array irradiated by an isotropic 11.6 keV photon point source. The analysis assumes a known emitted exposure. A dual-branch convolutional neural network combined the normalized 12 × 12 response with six global descriptors—the total photon rate, two centroid coordinates, two spatial spreads, and covariance—to estimate a front-hemisphere unit direction vector and the source-to-detector distance. The dataset contained 272 training, 96 validation, and 72 in-dependent test positions. Across three initializations, the full model achieved a polar-angle MAE of 5.16 ± 0.42°, circular azimuth MAE of 13.56 ± 0.85°, three-dimensional directional MAE of 9.68 ± 0.84°, distance MAE of 26.33 ± 1.43 mm, and MAPE of 6.29 ± 0.26%. Shape-only and Global-only ablations showed that the spatial pattern and global statistics contributed different information, and the full model outperformed Random Forest, Extra Trees, and template/k-nearest-neighbor baselines on the same test set. The dominant failure mode was azimuth estimation for far-field sources close to the array normal. These results are lim-ited to the simulated test positions and do not establish real-detector performance or reliable full-angle localization at 1 m.

An uncollimated scintillator array encodes source direction and distance in both its spatial response and total count rate. We modeled an isotropic 11.6 keV point source and a 12 × 12 yttrium orthosilicate array in Geant4 11.3.2, using a scintillation yield of 15,300 photons MeV⁻¹. A spatial–global dual-branch convolutional neural network, Full, combined normalized response patterns with six global descriptors for joint localization conditional on known emitted exposure. The response library comprised 272 training and 96 validation positions, with a 72-position multi-distance evaluation grid and 18 fixed-distance diagnostic positions. On the multi-distance grid, Full achieved three-seed mean directional and distance MAEs of 10.4° and 26.0 mm. At 300 randomly sampled positions and 4 million primaries, the corresponding means were 15.6° and 52.6 mm. An input-matched multilayer perceptron gave 16.7°/55.2 mm at 4 million primaries. Same-coordinate exposure controls and independent Monte Carlo repetitions characterized statistical sensitivity. A forward-response inverse estimator performed better in a far-field, near-axis subset. The results support combining spatial and rate information for localization under the simulated conditions and identify exposure and viewing geometry as major limits on accuracy.