Physics-Informed Deep Learning in Fusion Plasmas

The development of integrated models of fusion plasmas continues to be a challenge. Such models require the self-consistent treatment of a diverse range of physical phenomena whose integration is often computationally impractical. Deep learning methods offer the promise of drastically reducing the computational cost of describing several plasma physics processes thus facilitating the development of integrated models. This presentation will describe progress on the development of physics-informed machine learning descriptions of distinct plasma physics components. Each of these models is trained only on the underlying PDE, boundary conditions and initial conditions, without any experimental or simulation data, thus avoiding the need to generate large data sets. This presentation will describe progress on the development of surrogates for the relativistic Fokker-Planck equation describing runaway electron evolution in tokamak plasmas, the Vlasov-Fokker-Planck in the context of Knudsen layer reactivity reduction in inertial confinement fusion targets, the ion drift-kinetic equation in axisymmetric tokamak geometry, and the quasi-static MHD equations in the context of a vertical displacement event. The diversity of applications reviewed suggests the versatility of physics-constrained machine learning approaches to developing efficient physics modules for use in integrated simulations.