Advancing Ray-Traced Ultra-High Vacuum Simulations: Enhanced Algorithms and Data Structures in Molflow

Vacuum components play a critical role in many fields. Their design can be a tedious progress, as they are usually evaluated throughout many iterations with simulation software such as Molflow. Molflow uses the Test Particle Monte Carlo (TPMC) method to approximate physical quantities such as the pressure or particle density in arbitrary complex geometries. The TPMC traces independent particles inside a geometry using a ray tracing based algorithm to gather the corresponding statistics. The performance of these Monte Carlo simulations largely depends on the efficiency of the underlying data structure and algorithms used for the ray tracing query. As ray tracing kernels reached new heights leading to real-time rendering with recent advancements in both algorithm research and specialised GPU hardware, the utility of these state-of-the-art methods has been investigated and is evaluated for their suitablity for physical simulations such as Molflow. We provide enhanced and newly developed algorithms and data structure for Molflow's Monte Carlo model, in particular the time-dependent simulations. A specialised ray tracing kernel is developed based on our findings for both CPU-driven simulations leading to an application suitable for HPC environments, as well as GPU simulations where the design of a CUDA kernel is backed by NVIDIA's OptiX API for ray tracing to leverage hardware-acceleration utilising ray tracing units (RTUs) found on modern NVIDIA RTX GPUs. Further, we developed two splitting criteria to construct performant acceleration data structure, the adapted Ray Distribution Heuristic and the Hit Rate Heuristic. Both methods are designed to leverage Molflow's statistical nature. The developed solution for the GPU utilizes a newly developed offset to mitigate negative effects that arise from 32-bit floating point operations that are inherently used for hardware-accelerated ray tracing. Our GPU kernel utilizes RTUs as much as possible. The Neighbor Aware Offset handles some of the effects on the software side. Our research represents the initial step towards enabling Molflow simulations on GPUs.