OptiGAN for crystal arrays: physics-informed generative modeling of optical photon transport in PET detector arrays

Monte Carlo simulations of optical photon transport are computationally prohibitive for large-scale optical systems including detector arrays and PET systems, restricting their practical use to single-crystal studies. This work presents an enhanced conditional generative adversarial network (optiGAN) capable of replacing optical simulations at the crystal array level, extending our previous single-crystal approach to a 3×3 BGO detector array. We introduce Fourier feature encoding and a learnable latent mapping network as the modifications enabling stable training on the array geometry, together with a physics-informed loss term enforcing the unit-sphere state space S 2 of the generated propagation directions as a soft constraint. Training data requirements are reduced eight-fold by exploiting the array's symmetry. The model is evaluated in three studies: a full array evaluation, a high-resolution study probing generalization, and a pencil beam irradiation study assessing practical applicability. Performance is benchmarked against GATE10/Geant4 ground truth, using the fluctuations between independent Monte Carlo runs. The enhanced optiGAN achieves similarity values within 3σ-agreement of the Monte Carlo baseline across all evaluation conditions. An ablation and attribution analysis shows that the physics-informed loss term reduces low-SSIM bin fractions by a factor of 3.6 on the outer crystals, with a localized trade-off at the central crystal, yielding a net 48% reduction over the full array. The model transitions from electron-emission training data to realistic γ-photon interactions, producing flood maps that reproduce experimental patterns including photopeak clusters and inter-crystal scatter lines. This proof-of-concept demonstrates that a physics-informed generative model can simulate optical photon transport in segmented scintillator arrays at a training and inference cost accessible on a single workstation GPU. This provides a foundation for future models capable of generalizing across diverse array configurations.