Development of a multi-fidelity CFD framework for efficient parametric studies in nuclear thermal hydraulics

湍流的直接模拟通常是不可行的。一种更实用的方法是采用比最小尺度的湍流结构更粗糙的网格。然而，这会带来潜在的建模和离散化错误。我们提出了一种数据-科学/工程相结合的模拟方法，利用卷积神经网络对粗网格CFD(CG-CFD)进行“校正”；提供“最佳估计”和不确定性。与现有方法不同，我们的目标是能够对建模和离散化误差进行预测校正的特征向量集。我们建议在参数扫描的多保真框架中使用CG-CFD作为“低保真”组件。该框架将基于文献中现有的方法，使用深度高斯过程回归来桥接保真度。该项目得到了劳斯莱斯和IBM英国研究院的支持，他们表示通过直接参与这项工作进行实物捐助。由此产生的方法将主要在与核热液相关的测试案例上进行演示，但预计发展将与其他正在进行的研究领域相关，如流体-结构相互作用和多相流。

The direct simulation of turbulence is typically infeasible. A more pragmatic approach is to employ a mesh that is coarser than the smallest-scale turbulent structures. However, this introduces potential modelling and discretisation errors. We propose a combined data-science/engineering simulation approach, whereby Convolutional Neural Networks "correct" coarse-grid CFD (CG-CFD); providing the "best estimate" plus uncertainty. Unlike existing approaches, we target a feature vector set that will enable the prediction corrections for both modelling and discretisation errors. We propose to use CG-CFD as a "low-fidelity" component in a multi-fidelity framework for parametric sweeps. The framework will use Deep Gaussian Process Regression to bridge fidelities based on existing methods in the literature. The project has gained support of both Rolls Royce and IBM Research UK, who have indicated in-kind contribution via their direct involvement in the work. The resulting method will be demonstrated primarily on test cases relevant to Nuclear Thermal Hydraulics, but developments are expected to be relevant to other ongoing research areas such as fluid-structure interaction and multiphase flow.