From Music to Fusion: How artificial intelligence helps to interpret complex signals

7 mai 2024

Speaker :  Enrique Zapata-Cornejo (doctorant M2P2)

Abstract: Have you ever enjoyed listening to a particular instrument playing in a band? Differentiating musical waves is a natural thing for us. How can we teach machines to do it? Nowadays, you can talk with a machine so everything looks possible… Regarding other issues, the search for a clean alternative and unlimited energy source is underway: nuclear fusion, the reaction that fuels the stars, is a promising source of carbon-free energy. An ionized gas called plasma needs to be confined under specific conditions so atoms can fuse and release tremendous amounts of energy. However, the presence of certain waves in this plasma can represent an important obstacle to delivering this promise.
Under certain circumstances plasma waves can grow without control, destabilizing the plasma. Because of that, they are called instabilities. Can music inspire us to understand and detect these dangerous waves in plasmas? In our research, we have developed new approaches to automatic signal analysis, so we can identify these musical patterns in our fusion experiments. First, an algorithm inspired by music transcription is developed. A neural network looks for waves matching a code, much like musicians transcribe sounds to scores to play. We have used data from 1300 experiments and we have looked for patterns in their signals. Our second approach is slightly different; for years, experimental physicists looked at their computer screens to identify plasma instabilities. The researchers looked for very specific patterns in signal representations called spectrograms, which is an extremely tedious task. Therefore, we used computer vision and wavelet algorithms to recognize the same patterns, imitating this human behaviour and perform the analysis in a faster way.
This fusion research is an example of how signal analysis and AI research have broader implications, as the presence of the same data patterns can be present in many other disciplines.

The seminar take place on May 7th 2023 at 10am / amphi. N°3 / Centrale Méditerranée