Speaker
Description
The advent of artificial intelligence [AI] has a deep impact on numerous scientific and industrial fields, particularly in magnetic confinement fusion. This work explores the application of AI techniques to help scientists with the design of future fusion experiments based on previous experimental campaigns. Traditional ways of interpreting and designing fusion discharges often require extensive computational resources, time and research experience (including trial and error procedure). By leveraging AI, it is shown the possibility to partially overcome these constraints. The major goal of this work is the development of an AI system that is able to estimate the operation parameters given a desired plasma scenario. As a first prototype, the explored AI techniques are applied to the TJ-II stellarator. The produced operation parameters are the plasma fueling and heating configurations, which are coded into a set of 142 numbers that completely characterize the time evolution of the imposed ECRH, the NBI configuration and the puff-in plasma feeding. In this work, the plasma scenario is based on the magnetic fluctuations activity. It is measured by a Mirnov coil and its influence on the fast ion confinement is well known. The spectrogram from the Mirnov signal is coded by a Convolutional Auto-Encoder into a low-dimensional latent space. A multi-layer perceptron is afterwards trained and employed to link to spectrogram encoded in the latent space with the 142 numbers characterzing the discharge. The training is performed using a database composed of 2283 shots after data cleaning and filtering.
After training the AI model, it is possible to predict the machine configuration based on the desired Alfvén activity. The accuracy of the prediction is assessed by inspecting the 228 shots of the test set, finding that 70% present an excellent alignment with the ground truth, 15% approximately match the ground truth, whereas only 15% do not agree at all.
The results indicate that AI can approximate fusion experiments and assist scientists for the design of new ones, offering a faster and cost-effective alternative to conventional approaches. This study paves the way for more efficient research and development processes in fusion experiments, with AI serving as a tool for innovation and discovery. The next step for this work will be the inclusion of other plasma diagnostics in the system, such as the line density.
| Presentation type | Poster |
|---|
