Accelerating Tokamak Disruption Simulation with a Novel Framework for Runaway Electrons

One crucial challenge in achieving a successful tokamak design is the resilience to disruption events, either through prevention or mitigation. Despite large efforts over the past decades, the disruption problem is still unsolved, leading to existing and planned future machines providing a strong experimental testbed for designing disruption mitigation scenarios. A robust simulation framework, however, is lacking, due to the multi-physics nature of a disruption prohibiting first principles simulation. While reduced order models can be deployed to provide an integrated disruption model, they carry uncertainties that can lead to inaccuracies in predicting the damage from a particular disruption scenario, hence a surrogate model is needed to provide an efficient, but accurate description of a tokamak disruption. This motivation leads us to leverage modern scientific machine learning methods to provide such a tool and demonstrate a novel path forward in efficiently describing a disruption with high accuracy through physics-informed neural networks (PINNs) and an adjoint formalism to describe runaway electrons (REs) [1–4]; a key ingredient in accurately predicting the damage to plasma facing materials. The “DeepRunaway” framework is able to rapidly predict RE formation on the timescale of milliseconds, enabling direct coupling to the broader physics models of a disruption. This talk will provide an overview of the framework, specifically how various RE formation mechanisms can be accurately described, and demonstrate how a PINN can be deployed to learn a PDE in the presence of the large parameter space that emerge when exploring distinct disruption mitigation strategies. Together, a novel tool for predicting RE formation and an approach for leveraging PINNs for real world physics problems provide a promising path towards overcoming scenarios where traditional methods are computationally prohibitive.

[1] C. J. McDevitt, Physics of Plasmas 30 (2023).

[2] J. S. Arnaud, T. Mark, and C. J. McDevitt, Journal of Plasma Physics 90, 905900409 (2024).

[3] C. J. McDevitt, J. S. Arnaud, and X.-Z. Tang, Physics of Plasmas 32, 042503 (2025).

[4] J. S. Arnaud, X.-Z. Tang, and C. J. McDevitt, Submitted to Nucl. Fusion (arXiv:2504.03201) (2025).