GPU acceleration of optical photon propagation in low photon yield applications: Opticks for the Electron Ion Collider

TL;DR

The paper addresses the prohibitive cost of optical-photon transport in detector simulations by introducing EIC-Opticks, a GPU-accelerated fork of Opticks that uses event aggregation to batch thousands of Geant4 events into a single GPU run, keeping non-optical physics on the CPU. The approach yields physics-validated results against Geant4 with the pfRICH detector and delivers dramatic speedups, notably up to $161\pm3$× over single-thread CPU and around $3{,}000$× throughput gains with batching. Key contributions include guardrails for geometry and ray-tracing settings, a Spack package and containerized deployment for easy installation, and automated geometry-tuning utilities to optimize performance per geometry. This work enables fast, large-scale optical studies and ML-data generation for low-to-moderate photon-yield detectors and is readily generalizable to other Cherenkov and scintillator systems, including medical imaging.

Abstract

The bulk of time spent in the simulation of Cherenkov and other scintillation detectors is spent on optical-photon transport, i.e. ray tracing, a task that GPUs are uniquely qualified to perform. We present EIC-Opticks, a fork of Opticks, which uses event aggregation to drastically accelerate photon transport simulation for low-to-moderate photon yield experiments. During the full Geant4 Monte Carlo simulation of a given detector, optical photon simulation is performed on GPU(s) using the NVIDIA OptiX framework. We validate this approach using the ePIC pfRICH detector. We find GPU and CPU simulations in excellent agreement. For $5\times 10^4$ electrons with a momentum of $p=5~\mathrm{MeV}/c$ in the test case of the pfRICH detector, EIC-Opticks shows an order-of-magnitude speedup over multi-threaded Geant4, and a factor of up to 161$\pm$3 over single-threaded execution. In the case of low-to-moderate applications event aggregation reduces the per-photon simulation time from $\sim60\,μ\mathrm{s}$ for single events to $\sim20\,\mathrm{ns}$ with batching, a factor of $\sim3000$. In order to make EIC-Opticks easily installable, we authored a Spack package that makes it possible to install it with a single command. Additionally, a Docker container is provided for users with EIC-Opticks installed. EIC-Opticks provides guardrails for common pitfalls (e.g. nested volume conversion, ray tracing setting optimization).

Figures (9)

• Figure 1: The Geant4 simulation of a single 5 MeV electron (red track) entering the pfRICH detector, creating Cherenkov photons (green tracks) mainly in the aerogel tiles.
• Figure 2: 50,000 electrons with momentum 5 MeV/c shot in as a pencil beam. Hit maps simulated by Geant4 (CPU) validation on the left and EIC-Opticks (GPU) on the right. Event batching introduced in EIC-Opticks was utilized. The identical (within statistical errors) hit maps prove that EIC-Opticks yields the same results as Geant4.
• Figure 3: Speed increase of the GPU simulation compared to propagating the photons with multi-threaded Geant4 (CPU) for different number of Geant4 threads.
• Figure 4: Average time spent on the simulation of a single photon with EIC-Opticks for different event batch sizes in a single GPU call.
• Figure 5: Speed increase of the GPU simulation due to batching compared to no batching. For the pfRICH detector, simulating 10,000 events in a single GPU call yields a speedup of three orders of magnitude.