Accelerating High-Fidelity Parametric Thermal Quench Simulations via Neural Operator Preconditioning for Disruption Mitigation in Tokamaks

High-fidelity, nonlinear MHD simulations, such as those performed using the M3D-C1 code, are computationally intensive, limiting their practicality for extensive parametric studies and optimization tasks. In this work, we present a neural operator-based hybrid framework, Minimal Data Parametric Neural Operator Preconditioning (MD-PNOP), which is used to accelerate M3D-C1 simulations of impurity-driven thermal quenches. Accurate simulation of impurity-induced thermal quench dynamics across varying impurity profiles is crucial for understanding mitigation efficacy and optimizing relevant scenarios conditioning to the potential damage induced on the device [1]. Injection of impurities via mitigation systems to initiate controlled plasma termination is the chosen mitigation system for upcoming burning plasma devices [2]. Rapid radiative cooling of the plasma core is essential to dissipate thermal energy uniformly and protect plasma-facing components in particular in high-energy-density fusion devices such as SPARC. To accelerate M3D-C1, a neural operator [3] is trained on a minimal set of high-fidelity simulations and generalized to efficiently approximate solutions across a broad range of parameters. These approximate solutions are then refined and fully constrained using the physics-based M3D-C1 solver, ensuring consistency with the underlying plasma dynamics and maintaining high accuracy. This approach aims to significantly reduce computational cost while preserving fidelity, enabling systematic exploration of impurity scenarios and supporting the development of robust disruption mitigation strategies in next-generation high-field fusion devices.



[1] C.F. Clauser, and et. al., Modeling of carbon pellets disruption mitigation in an NSTX-U plasma, Nuclear Fusion, 61 116003 (2021).

[2] E. M. Hollmann, and et. al., Status of research toward the ITER disruption mitigation system, Phys. Plasmas, 22, 021802 (2015).

[3] K. Azizzadenesheli, and et. al., Neural operators for accelerating scientific simulations and design, Nat. Rev. Phys, 6, 320-328 (2024).