Generative Models and Simulation to Assess Uncertainties for Tokamak Infrared Thermography

Quantitative infrared thermography, which aims to infer surface temperatures from an infrared image, is challenging in metallic environments. A local increase in the flux measured by the camera may be due to indiscriminately an increase in the temperature of the observed target or due to a change in its surface properties. This paper addresses this ill-posed inverse problem and provides a method to automatically evaluate a confidence interval on the predicted temperature. For this purpose, we use simulated data, generated with ray tracing, to train a deep generative model able to approx-imate an unknown probability distribution. We apply the proposed method to simulated infrared images of fusion fa-cilities, where infrared thermography is a critical diagnostic for machine protection. We show that the model estimates the temperature with a 6% mean error, and provides a 95% confidence interval, on average 20% of the predicted value, improving the reliability of the measurement.