A Data-Driven Multi-Fidelity Framework for Enhanced Flow Prediction Around Propeller and Fan Tips

由于固有的假设和经验主义，低保真度的湍流模拟模型在涉及分离、再循环和再附着的复杂流动中缺乏所需的精度。特别是，螺旋桨和转子尖端周围的湍流漩涡流动对大多数低保真螺旋桨模型提出了重大挑战。另一方面，高保真尺度解析方法，如大涡模拟（LES）或延迟分离涡模拟（DDES），对于设计优化和不确定性量化等多查询计算任务来说，计算量太大。该项目旨在开发一种新颖的、数据驱动的多保真度螺旋桨建模框架，该框架由最先进的基于伴随的流动同化技术和机器学习（ML）模型组成，具有增强的螺旋桨和转子尖端周围流动的预测能力。我们提出通过引入校正项来增强低保真螺旋桨模型。采用反设计方法，对低保真螺旋桨模型的校正项进行了调整，以减小增强模型与给定高保真样本之间的差异。在流动同化过程中计算的平均流动特征和校正项之间的输入输出对用于训练嵌入求解器中的卷积神经网络，以增强其对螺旋桨和风扇尖端区域周围复杂流动的预测，而无需依靠计算密集型的尺度分辨模拟。在本研究中，除了利用高保真度模拟来研究流动物理和噪声产生机制外，我们还提出利用高保真度数据，通过迁移学习和主动学习，迭代提高基于低保真度数据构建的基于ml的螺旋桨噪声模型的预测精度，从而实现一个多保真度数据驱动的螺旋桨噪声预测和最小化框架。总而言之，该项目的目的是利用先进的机器学习技术开发一个多保真度的螺旋桨性能预测框架，从而为行业迫切需要的高效预测和设计高效低噪音螺旋桨和风扇提供关键支持技术。在EPSRC的景观中，该项目与工程研究主题中的流体动力学和空气动力学研究领域非常吻合。它也与EPSRC人工智能和机器人主题领域的关键战略投资之一“科学和政府人工智能”计划保持一致。

Due to the inherent assumptions and empiricism low-fidelity models for turbulent flow simulation are well known to lack the desired accuracy in complex flows involving separation, recirculation and reattachment. In particular, turbulent swirling flows around the propeller and rotor tips present significant challenges to most low-fidelity propeller models. On the other hand, high-fidelity scale-resolving methods such as the Large Eddy Simulation (LES) or Delayed Detached Eddy Simulation (DDES) are too computationally intensive for many-query computational tasks such as design optimization and uncertainty quantification. The proposed project aims to develop a novel, data-driven multi-fidelity framework for propeller modelling, consisting of a state-of-the-art adjoint-based flow assimilation technique and a machine learning (ML) model, with enhanced predictive capabilities for flows around propeller and rotor tips. We propose to enhance a low-fidelity propeller model by introducing correction terms. Using inverse design, the correction terms to the low-fidelity propeller model are tuned to reduce the discrepancies between the augmented model and a given high-fidelity sample. The input-output pairs between the mean flow features and correction terms computed in the flow assimilation process is used to train a convolutional neural network that is embedded in the solver to enhance its prediction of complex flows around propeller and fan tip regions without resorting to the computationally intensive scale-resolving simulations. In the proposed work, in addition to using high-fidelity simulations to study the flow physics and noise generation mechanisms, we propose to leverage the high-fidelity data to iteratively enhance the predictive accuracy of the ML-based propeller noise model constructed based on low-fidelity data, via transfer learning and active learning, and in so doing, realize a multi-fidelity data-driven framework for propeller noise prediction and minimization. In summary, the aim of this project is to develop a multi-fidelity framework for propeller performance prediction using advanced machine learning techniques, thus providing a key enabling technology the industry urgently need to efficiently predict and design high-efficiency low-noise propellers and fans. In the EPSRC landscape, this project aligns well with the research area of Fluid Dynamics and Aerodynamics within the research theme of Engineering. It is also well aligned with the 'AI for Science and Government' programme - one of the key strategic investments within the EPSRC Thematic Area of AI and Robotics.