Physics informed deep learning for fusion thermal hydraulics

磁约束聚变的一个重大挑战是面向等离子体的组件的高热负荷。高能中子轰击这些组件，产生不均匀的体积热负荷，必须转移以防止热损害。虽然数值模拟为这些部件的设计提供了一条诱人的途径，但其高昂的成本、广度和高维限制了设计空间探索的有效性。在本项目中，将研究物理感知的深度学习代理模型作为全流体PDE建模的候选模型的潜力。我们的动机是有效地探索熔融热工水力学的参数空间。物理信息神经网络(PINN)技术已经在无磁流体(MHD)效应的等温流动中显示出巨大的前景，但还需要进一步发展以考虑与聚变相关的条件。PINN将用于广泛和快速的参数空间探索，而选定的高阶计算流体动力学(CFD)研究(作为本PHD的一部分)将用于模型的训练和测试。这项工作将集中在湍流浮力MHD，具有共轭换热和非均匀体积加热。案例研究是加热腔和蛇形通道结构，这两种结构都是聚变中包层/偏滤器设计的中心。这将对面向等离子体的部件中存在的复杂流动物理提供至关重要的理解，并有望导致未来聚变工厂的改进设计。

A significant challenge in magnetic confinement fusion is the high heat loading of the plasma-facingcomponents. High energy neutrons bombard these components, creating a non-uniform volumetric heatloading which must be transferred to prevent thermal damage. While numerical simulation offers anattractive route forward for the design of these components, its high cost, breadth and highdimensionality limits efficacy in design space exploration.In this project, the potential of physics aware deep learning surrogate models as candidates for full fluidflow PDE modelling will be investigated. Our motivation is efficient exploration of parameter space forfusion thermal hydraulics. Physics informed neural network (PINN) techniques have already shownsignificant promise in isothermal flows without magnetohydrodynamic (MHD) effects but need furtherdeveloping to account for fusion relevant conditions. PINNs will be used for broad and rapid parameterspace exploration while selected high-order Computational Fluid dynamics (CFD) studies (generated aspart of this PhD) will be used in training and testing of the model.This work will focus on turbulent buoyant MHD, with conjugate heat transfer and inhomogeneousvolumetric heating. The cases studies are a heated cavity and serpentine passage configuration, bothbeing central to blanket / divertor design in fusion. This will provide a vital understanding of the complexflow physics present in the plasma-facing components and is expected to lead to improved designs infuture fusion plants.