High fidelity micro- and meso-scale computations and data-driven physics-informed models of particle-laden flows

微粒流在自然界和工业中无处不在，从地球和海洋科学中的河流沉积物运输，到工艺工程中的流化床化学反应器，再到生物医学工程中人体静脉和毛细血管血液系统中的药物运输。虽然已经有大量的关于颗粒流动力学的知识，但主要的相间动量传递的复杂性是研究人员和工程师仍然无法完全理解这类多相流的主要原因。当悬浮液不能再被稀释时（这是许多应用中常见的情况），颗粒强烈干扰邻近颗粒周围的流场，导致大量颗粒间的水动力和扭矩波动，其幅度往往与平均值相当。颗粒流动的大尺度数值模型，如中尺度欧拉-拉格朗日（EL）模型和宏观尺度欧拉-欧拉（EE）模型，依赖于相间动量和热量传递的闭合律。目前的闭合定律无法预测重要的粒子间波动，而这是EL和EE模拟保真度的关键。由于EL和EE模型具有实际和工业用途，因此提高其仿真的保真度非常重要。该项目主要关注EL模型。高保真微尺度粒子分辨模拟（PRS）提供详细的信息，不需要任何封闭，但需要非常大的计算资源，只能模拟小型系统。本项目的目标是设计全新的相间传递模型，预测平均值和粒子间波动。为此，我们分析了微尺度PRS产生的大型数据集，并将我们从该分析中学到的知识以数据驱动物理信息（DDPI）模型的形式转移到中尺度EL模型中。虽然使用传统方法分析大型PRS数据集仍然是一个有价值的研究路径，但我们使用机器学习和神经网络（NN）从数据中推断出额外的理解，并设计更先进的间期转移模型。这种方法代表了这一领域范式的彻底改变。DDPI模型代表了颗粒流的新一代相间传递模型和以前无法达到的目标保真度水平。作为混合模型，它们汇集了两个世界的精华：物理理解和机器学习。自然界和工业中无处不在的微粒流支持了持续研究的必要性。我们特别感兴趣的应用领域是过程强化和绿色能源生产，包括减少这些过程的环境足迹，以及开发涉及木材和阳光等可再生材料的可靠和有效的新技术。

Particle--laden flows are ubiquitous in nature and industry, ranging from sediment transport in rivers in earth and ocean science, fluidized bed chemical reactors in process engineering to drug transport in the human blood system of veins and capillaries in biomedical engineering. While a large body of knowledge already exists on particle-laden flow dynamics, the complexity of the dominant interphase momentum transfer is the primary reason why a complete understanding of this class of multiphase flow still escapes researchers and engineers. When the suspension cannot be regarded as dilute anymore, a common situation in many applications, particles strongly disturb the flow field around neighboring particles, leading to substantial particle--to--particle hydrodynamic force and torque fluctuations of magnitude often comparable to the mean value. Large-scale numerical models of particle-laden flows, such as the meso-scale Euler-Lagrange (EL) model and the macro-scale Euler-Euler (EE) model, rely on closure laws for the interphase momentum and heat transfer. The current closure laws are incapable of predicting the significant particle--to--particle fluctuations that are key to the fidelity of EL and EE simulations. Improving the fidelity of EL and EE simulations is of tremendous importance as these two models are of practical and industrial use. The project focuses on EL models. High fidelity micro--scale Particle-Resolved Simulation (PRS) supplies detailed information without the need for any closure but require very large computing resources and can simulate small systems only. The objective of this project is to design fully novel interphase transfer models that predict both the mean value and the particle-to-particle fluctuations. To do so, we analyze large data sets produced by micro-scale PRS and transfer the knowledge we learn from this analysis to the meso-scale EL models in the form of Data--Driven Physics--Informed (DDPI) models of interphase transfer. While the analysis of large PRS data sets with traditional methods remains a valuable research path, we use machine learning and neural networks (NN) to infer additional understanding from the data and to design more advanced interphase transfer models. This approach represents a complete change of paradigm in this field. DDPI models represent the next generation of interphase transfer models in particle--laden flows and target fidelity levels that were previously unattainable. As hybrid models, they bring together the best of two worlds: physical understanding and machine learning. The omnipresence of particle--laden flows in nature and industry supports the need for sustained research. Fields of application of particular interest to us are process intensification and green energy production, including the reduction of the environmental footprint of these processes and the development of reliable and efficient new technologies involving renewable materials such as wood and sunlight.