Speaker
Description
Plasma turbulence and the resulting transport of particles and energy remain among the most consequential, elusive problems in fusion science. In magnetic-confinement fusion (MCF) devices, nonlinear, multiscale instabilities and turbulence arise across various plasma regions that strongly influence plasma confinement, pedestal structure, transient events such as edge-localized modes, and power exhaust to the divertor. These processes ultimately determine whether MCF devices can sustain energy confinement long enough for practical fusion reactor performance while protecting plasma-facing wall components.
Our knowledge of several aspects of the turbulence-induced transport problem – for instance spatiotemporal evolution of turbulence and wave-particle interactions driving transport – remains incomplete. This gap can be largely attributed to the lack of predictive, cost-efficient models, as the highest-fidelity simulations are prohibitively costly at reactor-realistic scales and for operational timeframes, while lower-computational-cost, fluid-based models lack generalizable closure relations to capture essential kinetic physics they cannot resolve.
In this landscape, ML has potential to address the gap. However, the approaches pursued so far typically face one or more of the following limitations: (i) soft penalty enforcement of physics constraints, leading to model drift from physics principles; (ii) reliance on expert-defined candidate-term libraries and/or explicit physical-grounding optimization constraints that may be incomplete or biased; (iii) challenges in generalizing across regimes, geometries, and operating points; and (iv) black-box behavior with limited interpretability. Together, these issues explain why existing ML tools have struggled to deliver models that are suitable for scientific discovery in fusion science.
This work introduces neuro-symbolic AI as an alternative to address the outstanding scientific questions related to turbulence-induced transport in nuclear fusion plasmas. Neuro-symbolic AI combines the power of neural networks to extract (learn) key dynamics from high-dimensional data with the capacity of symbolic AI to enforce fundamental physics constraints – such as symmetry, invariance, and conservation laws – via logic-based reasoning, guiding the equation discovery process within physically admissible solution spaces. The paper will elaborate on the unique ways through which neuro-symbolic AI can suit fusion science problems and discuss the challenges that must be addressed to realize the full potential of the approach on path to harness nuclear fusion energy.