Speaker
Description
In fusion devices, infrared thermography is extensively used to monitor surface temperatures ranging from 100 to 3600°C. However, interpreting infrared measurements in reflective metallic environments, such as those found in fusion devices, presents significant challenges. These complexities often result in temperature measurement errors on critical components, posing operational risks. To address these issues, we have developed an infrared image simulation tool based on the Monte Carlo Ray Tracing method. This tool is capable of simulating intricate thermal scenes, including detailed 3D geometries characterized by complex optical properties. Nevertheless, the primary limitation of the Monte Carlo method is the extensive sampling required to obtain usable images. This can escalate to billions of rays for a single image of 512x640 pixels, inevitably increasing computational costs.
Our recent advancements have focused on accelerating the code by leveraging NVIDIA GPU's RT-core technology to enhance efficiency and reduce computational time. This technology able to compute in hardware the ray-triangle intersection as well as the Bounding Volume Hierarchy (BVH) acceleration structure traversal, which involves dividing the scene into multiple sub-regions, creating a tree-like graph. This allows rays to traverse the structure and interact only with relevant triangles in the 3D scene, rather than comparing with all triangles, both significantly optimizing the ray-tracing process. When applied to a thermal scene in fusion devices comprising 1 million triangles, the simulation of a 512x640 pixel infrared image with 1 billion rays shows a remarkable improvement of 450 times. The simulation now takes just 5 seconds on an NVIDIA A40 GPU, compared to 32 minutes on a 48-thread Intel processor. Such performance is crucial for generating large synthetic databases used for training machine learning algorithms.
This new simulation has been integrated into CIVA, a widely-used software for simulation and analysis in Non-Destructive Testing (NDT), both in industry and research. This integration enables near real-time analysis of experimental data by varying different parameters such as temperature, emissivity, reflection models, and camera settings. It also facilitates the discrimination of false hot spots caused by reflection patterns or the underestimation of temperature due to changes in emissivity, which can pose operational risks. In further applications, these simulations have been used to build inverse models for estimating thermal scene parameters, such as temperature and emissivity, from infrared images.
