托卡马克装置是实现可控核聚变的重要途径，装置中等离子体撕裂模不稳定性的发展及饱和过程的理解对维持托卡马克稳态运行有重要意义。受制于计算资源，传统MHD数值求解器保真度普遍不高：基于显式或半隐式算法、采用低精度算法、计算规模也比较有限。异构众核构架使得HPC平台走向E级计算，也使高保真求解磁流体动力学方程成为可行，但也因访存、通信、能耗等限制为大规模仿真程序的编写带来了新挑战。. 本项目拟在前期研究的基础上，探索面向异构众核构架的高保真磁流体动力学求解算法：以众核环境下的区域分解算法为基础，实现多场全隐高精度计算；利用人工神经网络算法对求解过程优化，通过提高计算访存比、计算通信重叠和众核协同策略等探索在异构众核平台上实现超大规模磁流体仿真的新思路；实现对真实托卡马克位形下等离子体中撕裂模形成、发展和饱和的机理研究，提高托卡马克等离子体的约束水平。

The Tokamak device is one which can effectively achieve controllable nuclear fusion power generator. To guarantee a stable operation of Tokamak, it is necessary to numerically analyze the real physical Tokamak device in terms of the potential of instability arising from tearing modes of plasma and saturation processes. However, due to the limitation of available computing resources, magnetic fluid solvers are traditionally developed based on explicit formats which are defective in terms of long time steps or non-linear solver stability. Fortunately, with the development of high-performance computing platforms, the large scale computing platform has enabled the possibility to solve the magnetohydrodynamic equations in full implicit format. Simultaneously, heterogeneous computing has become the mainstream architecture of large-scale parallel platforms which has resulted in a major challenge to the parallel scalability of the solver algorithms. .Therefore, beginning from the previous research on such algorithms, this project shall combine the parallel domain decomposition algorithm and artificial neural network to develop a highly precise and fully implicit magnetohydrodynamics solver for a large-scale heterogeneous computing platforms. In doing so, the mechanisms behind the instability arising from tearing modes of plasma and other issues in the Tokamak device can be deeply studied. Moreover, the prospect of high-fidelity magnetohydrodynamics numerical simulations using a large-scale parallel platform is further explored and a breakthrough of the bottleneck problem of numerical simulation to tokamak nuclear fusion can be expected.