U-Net for temperature estimation from simulated infrared images in tokamaks

Abstract

Surface temperature measurement is critical for the safety and proper operation of nuclear fusion reactors such as tokamaks, which operate at high temperature [100–3600 °C] in the presence of complex and unknown surface reflectance properties, causing a reflected flux that disturbs measurement interpretations. This paper describes a numerical approach based on machine learning techniques to estimate the surface temperature of in-vessel components from infrared images, despite the presence of reflections and the unknown emissivity of materials. Our contributions are two-fold: for the first time in the infrared domain, we generate a huge dataset of simulated images thanks to a novel GPU implementation of a Monte Carlo infrared ray tracer. This implementation offers the great advantage of enabling the generation of a very large amount of synthetic images for different plasma scenarios. Hence the second contribution of this paper: we propose an infrared inverse lighting model, based on a Convolutional Neural Network (CNN) trained with the images generated by our ray tracer. The proposed approach is applied on a numerical prototype of W Environment Steady state Tokamak (WEST), including complex simulated thermal scenes. Using synthetic test data, and without providing the surface optical properties, the mean temperature error predicted by the CNN on ITER-like wide angle view of WEST is evaluated to 9% with standard plasma scenarios and 22% in case of exotic thermal scenes.