- author = "Andreas Gillgren",
- title = "Interpretable machine learning applied to fusion plasmas",
- school = "Department of Physics and Astronomy, Chalmers University of Technology",
- year = "2026",
- address = "Gothenburg, Sweden",
- keywords = "genetic algorithms, genetic programming, JET, Magnetic confinement fusion, Machine learning, Artificial intelligence, Interpretability, XAI, Tokamak, Pedestal",
- isbn13 = "978-91-8103-424-0",
-
URL = "
https://research.chalmers.se/publication/551933/file/551933_Fulltext.pdf",
-
DOI = "
10.63959/chalmers.dt/5881",
- size = "108 pages",
-
abstract = "Magnetic confinement fusion is a field of research
that strives to develop an environmental friendly
energy source to assist in powering our society. By
confining a plasma with magnetic fields, conditions
that enable nuclear fusion can be achieved. However,
gaining a high efficiency has proven to be a
challenging task, which is why fusion research is still
at the experimental stage. On another scientific front,
machine learning (ML) and artificial intelligence (AI)
are advancing with tremendous momentum, and are rapidly
being integrated into fusion research. However, the
black-box aspect of many popular ML architectures
inhibits our ability to interpret what the models have
learned during training. Ideally, we would like to
understand ML models deployed in fusion research to
gain knowledge that may prove useful for advancing the
field.
This thesis focuses on exploring interpretable ML methods in fusion research. As a consequence, an interpretable framework called NeuralBranch has been developed, which has been applied to two different use cases in fusion. The main application in this thesis relates to the so-called pedestal, which has significance for the energy confinement in fusion experiments. The other, more secondary application in this thesis, relates to the growth rate of plasma instabilities that contribute to heat and particle transport. In summary, the interpretability of the machine learning models deployed reveals intricate parameter relationships in both these applications, beyond what previous traditional data-fitting approaches have been able to reveal.",
-
notes = "Mention of GP
Syupervisors: Par Strand and Dmytro Yadykin",
