Speaker
Description
The measurement of the total radiative power emitted by the plasma is a key aspect of diagnostics in tokamak fusion reactors, as this quantity represents the energy lost through electromagnetic radiation and is crucial for managing the energy balance. For these measurements, bolometric diagnostics are widely used in fusion experiments due to their precision, reliability, and ease of calibration. These detectors work by converting incident radiation into heat, causing a temperature increase in a thin absorber (typically a gold foil), which induces a change of the electrical resistance in a Wheatstone bridge. However, their time resolution is limited to a few milliseconds, while many plasma events occur on much shorter timescales. To address this, AXUV diodes, p-n junction photodetectors sensitive to visible light, UV, and soft X-rays, are also employed. Their high temporal resolution enables the detection of fast plasma phenomena such as instabilities and confinement losses. The disadvantages are that it has not been possible so far to implement an absolute in-situ calibration of this type of detectors and to extend their spectral response to the visible range..
In this work, a machine learning-based approach is proposed to predict bolometric measurements from AXUV diode signals, effectively enabling the reconstruction of plasma emissivity with high temporal resolution. Neural networks have been trained to learn the mapping between diode measurements and the corresponding output of the bolometric system. This data-driven model exploits the complementary nature of the two diagnostics systems, capturing the underlying correlations despite their differing spectral sensitivities.
Particular attention is devoted to autoencoder type neural networks, which are able to learn a compressed and meaningful features from the high-dimensional diode signals. Once trained, the model can provide synthetic bolometric signals at high temporal resolution, allowing the study of radiative behaviour during fast transient events that are otherwise inaccessible with conventional bolometric diagnostics.
