Hardware ray-traced optical-photon Monte Carlo for radiation detector simulation: a TOPAS extension on Apple Silicon

Optical-photon Monte Carlo (MC) underlies scintillator dosimetry, Cherenkov imaging and other medical-physics applications, yet Geant4 transport of the 10⁵-10⁶optical photons per event on the central processing unit (CPU) makes dataset-scale studies impractical even on a multi-core workstation. We implement and verify a hardware ray-tracing optical-photon transport extension for TOPAS on Apple Silicon. Approach. Charged-particle transport remains in Geant4 on the CPU, where the extension computes the scintillation and Cherenkov yields. Optical-photon generation, propagation and scoring run on the graphics processing unit (GPU), which finds boundaries with its ray-tracing hardware or, for curved solids, by solving the surface equation. The optical processes follow their Geant4 counterparts, with the dielectric-dielectric boundary physics algebraically identical. We verify the extension with ten benchmarks (B1-B10), covering optical physics, detector-chain timing and curved-surface intersection, on an Apple M3 Ultra with an 80-core GPU and a 30-thread CPU baseline. Main results. Every verification endpoint passes the equivalence test declared in advance, with margins of ±1 % in fluence and ±1 percentage point in the reflection and transmission probabilities. The largest mean summed-fluence deviation is 0.147 %, and every summed-fluence 95 % confidence interval lies within ±1 %. The scintillation-reflector-wavelength-shifting chain agrees to -0.02 % and the curved-geometry imaging chain to -0.07 %, and the intersection study reproduces the geometric chords of whole solids within a 0.5 mm tolerance. The matched scintillator-lens-camera simulation runs 15.7-32.0 times faster in execution time and about 24 times faster end to end at 10⁵events. Significance. The extension reproduces the analytical or CPU-Geant4 reference process by process within these margins, while accelerating matched transport by more than an order of magnitude. TOPAS scripts gain GPU acceleration without user-written C++, enabling dataset-scale generation of scintillation-dosimetry training sets and Cherenkov-imaging forward models on a single workstation. .