A physics-constrained deep learning surrogate of the runaway electron avalanche

The presence of runaway electrons (RE) during a tokamak disruption continues to motivate development of high fidelity descriptions that can be rapidly used for scenario optimization. One of the primary challenges in developing an integrated description of a tokamak disruption is the coupling of distinct physical components. The present work uses a physics-informed neural network (PINN) to derive a surrogate model of the RE avalanche, a process that is thought to be the dominant means through which REs are generated during reactor scenarios. In contrast to traditional deep learning approaches, PINNs do not require data and are instead constrained by the governing equations themselves. In this work we utilize a PINN to identify the parametric solution to the adjoint of the relativistic Fokker-Planck equation. The resulting PINN is used to evaluate the RE avalanche growth rate across a range of plasma conditions, thus providing an efficient surrogate of this process. Ongoing work is focused on extending the model to include partially ionized impurities and distinct large-angle collision operators, thus enabling the development of a high physics fidelity, but efficient description of RE generation in tokamak plasmas. [1] J.S. Arnaud et al., JPP, in press (arXiv:2403.04948) (2024)