A machine learning case study in nuclear fusion: Assessment of the absolute deuterium-tritium fusion power of ITER with gamma-ray spectroscopy

Abstract

Nuclear fusion holds great potential as a carbon-neutral means of electricity production. However, technical aspects of its implementation remain challenging. The real-time measurement of the fusion power released during Deuterium-Tritium (DT) fusion is one such aspect. The use of tools from artificial intelligence may help to solve this issue.

Recently, during experiments performed at the Joint European Torus, a novel method was developed to measure the fusion power in magnetic confinement fusion devices. Said method exploits the fact that gamma-rays released by the DT fusion reaction can be registered with a gamma-ray spectrometer. Expanding on this work, a machine learning algorithm was developed to estimate DT fusion power at ITER by use of the Radial Gamma-Ray Spectrometer (RGRS) measurements, as well as the magnetic equilibrium as an additional source of information.

The algorithm was trained and tested on a set of 75 simulations of ITER DT plasma scenarios. By testing the algorithm by repeated 5-fold cross-validation, the average deviation of the estimated fusion power from the reference was found to be 0.32%, while the relative error had a standard deviation of 0.97%. When statistical fluctuations were included in the analysis, the lowest measurable fusion power resulted to be around 30 MW, making the RGRS suitable for the fusion power measurement requirements at ITER.

This project demonstrated that a machine learning approach leads to promising results when coupled with prior knowledge and the integration of various kinds of sensor and simulation data. This and related algorithms may eventually contribute to the development of fusion power as a reliable, carbon-neutral source of energy.