Interpretable Machine Learning Accelerating Fusion Research

It has become widely accepted that Machine Learning (ML) accelerated research can enable reactor-relevant solutions for a broad spectrum of fusion challenges. Both inertial and magnetic confinement fusion need to address complex multi-scale, multi-physics systems whose integrated modeling implies extremely expensive computations, and ML can assist via surrogate modeling for accelerating such demanding simulation loops [Rodriguez-Fernandez 2022 NF 62 076036, Humbird 2021 PoP 28 042709]. Further relevant examples of ML applications in fusion include its adoption to enhance the analysis of instrumentation data [Samuell 2021 RSI 92 043520], to optimize experimental design and performances [Gopalaswamy 2019 Nature 565 581, Humphreys 2020 JFE 39 123–55], and for real-time monitoring of the proximity to different boundaries of plasma stability [Rea 2021 IAEA EX/P1–25].

This tutorial will start with a general description of Artificial Intelligence, and then focus on the specifics of Machine Learning and Deep Learning paradigms. Particular attention will be given to reviewing ML techniques that guarantee an explainable and interpretable predictive output, thus enabling effective controllers for magnetically confined fusion plasmas [Barr 2021 NF 61 126019]. Transfer learning and domain adaptation will also be discussed, since a common need exists to understand how to extrapolate knowledge to devices yet to be built or to experiments with different statistical properties [Zhu 2021 Nucl. Fusion 61 114005, Gaffney 2021 PoP 26 082704]. Finally, several examples of data-driven fusion applications will be provided, with particular emphasis given to active ML research conducted at the DIII-D facility.

This work is supported by the U.S. DOE under Award(s) DE-FC02-99ER54512, DE-SC0014264, DE-SC0010720, DE-SC0010492, and DE-FC02-04ER54698.