Speaker
Description
This work presents a new implementation of the radiative exchange factors calculations for 3D geometry using Graphics Processing Units (GPUs) with hardware-accelerated ray tracing capabilities.
Radiative exchange factors (REFs) represent the overall exchange of thermal radiation energy between two surfaces. Although there are analytical solutions for particular geometric configurations, in practice REFs are computed using numerical methods. One common approach involves employing Monte Carlo ray tracing algorithms. In a Monte Carlo ray tracing process, the electromagnetic energy emitted by the surfaces is divided into discrete rays with an associated energy. Then, each ray is traced from the surface it is emitted to the surface it is finally absorbed, considering any interaction it may have with the rest of the surfaces of the geometry. These interactions include specular and diffuse reflections and transmissions. However, a large number of rays need to be traced to achieve precise results, making this process highly computationally expensive. Because it is assumed that rays do not interact between each other, the calculation of each ray is independent from the rest, and the computation can be easily parallelized. Because of this, Monte Carlo ray tracing algorithms are very suitable for GPUs, especially the new ones with built-in hardware compute units to accelerate the process of tracing rays within the geometry.
In order to better utilize the GPU hardware capabilities, the geometry surfaces are discretized into triangular meshes. While this approach introduces a triangulation error to the result, the error can be reduced by refining the triangular meshes. Using Vulkan, an open-source, multi-platform standard for low-level access to the GPU hardware, we have developed a code focused on the architecture of the latest GPUs to calculate view factors (VFs) and REFs of STEP-TAS geometry models.
The results show that the GPU code we have developed can compute VFs and REFS with a negligible triangulation error much faster than other commercial software that only uses the CPU. As an example, using a single desktop GPU, we have been able to calculate the VFs and REFs of the Vigil’s PMI instrument Filtergraph model between two to three orders of magnitude faster.
