A Machine Learning based approach for Accelerating Z-Pinch Fusion Device Optimization

The shear flow-stabilized Z-pinch device is a popular approach to achieving controlled fusion energy. Designing and optimizing a Z-pinch reactor to maximize fusion performance requires exploring a vast, multi-dimensional space of device parameters. The most accurate predictive tools for this task are kinetic plasma simulations, which are capable of resolving non-linear phenomena at microscopic scales that are not captured in fluid models. However, the computational expense of these codes for the general case of a 6-D time-dependent PDE makes them impractical for the thousands of evaluations needed for systematic optimization. The study of using machine learning based surrogate models proves to be a promising area of research in simulating plasmas [G. Dong et al 2021 Nucl. Fusion]. The adopted approach involves training a deep neural network to learn the mapping from initial conditions (density, temperature and velocity profiles) to performance indicators such as time-evolving current flowing through the Z-pinch, which ultimately influences fusion yield. The training data for such a model is generated by solving the Vlasov-Poisson-BGK set of equations for a carefully selected set of initial conditions. Once successfully trained, the surrogate modeling approach has the potential to make the optimization problem of maximizing the time-integrated neutron yield tractable and computationally efficient in higher dimensions.