IMT: Development of LES and RANS models for H2 turbulentcombustion leveraging DNS data

与经典流体力学和动力学相比，计算流体动力学(CFD)在科学界是一个相对较新的工具。现代硬件的计算能力不断增强，使得像Navier Stokes这样的方程使用迭代求解器求解，即使是一台笔记本电脑也很常见。然而，许多流动现象仍然很难精确建模，湍流就是其中之一[1]。目前已有许多相对成功的模型使用雷诺平均纳维斯托克斯(RANS)和大涡模拟(LES)来模拟湍流，这就是为什么这些技术中的许多技术被用于显著加快从飞机到汽车的工业环境中的设计过程。燃烧和火焰传播是CFD有望成为未来发展基石的另一个领域。随着向可持续发展的明显转变，氢燃烧因其不产生二氧化碳的优势而处于能源生产和推进的前沿[2]。不幸的是，尽管氢是最简单的分子，但它的燃烧过程绝非如此。它受到热扩散不稳定性的强烈影响，当与湍流耦合时，热扩散不稳定性在火焰前锋传播中产生一系列非线性[3]，[4]。了解这些非线性是至关重要的，以便尽可能高效地在燃气轮机(GT)和内燃机(ICE)中使用氢气，以最大化其能量输出，并确保它们的安全运行。这是必要的，因为在给定体积的氢气中存储的能量比化石燃料中存储的能量要少得多，这使得效率成为设计者的首要任务。热扩散不稳定的复杂三维现象及其与湍流的耦合，导致火焰速度增加和火焰起皱，严重依赖于局部反应速率的变化。这意味着，在湍流区域，火焰是高度不规则的，这是由于湍流和氢的热扩散不稳定性的共同作用。这就产生了舌状结构，穿透在未燃烧的气体区域，这是甲烷等传统燃料的湍流燃烧中不存在的。这种高度不规则的流动很难成功建模，这导致了本项目的目标。因此，人们设想建立这些现象的准确模型，以帮助工业设计过程。直接数值模拟(DNS)的计算非常昂贵，除了学术环境外，不能被认为在任何地方都是可行的。因此，通过使用现有的DNS数据并进行进一步的模拟，计划产生可用于LES甚至RAN的低阶模型。此外，除了基于显式方程[5]的经典建模方法外，还可以使用机器学习(ML)[6]。ML是一个强大的工具，它允许绘制来自数据集的参数之间的关系。因此，可以通过基于DNS数据训练算法来创建上述模型。这将利用大学现有的高性能计算(HPC)硬件，然后可用于根据初始数据集或其他现有的LES和RAN模型重新测试生成的模型。因此，目标是在项目结束前产生更新的甚至新的氢气湍流燃烧模型，将为在更复杂的环境和模拟中应用这些模型铺平道路，并根据实验数据验证它们的性能，从而为在工业应用中实施氢气燃烧迈出第一步。

Computational Fluid Dynamics (CFD) is a relatively new tool in the scientific community compared to classical fluid mechanics and dynamics. The continuously increasing computational capabilities of modern hardware, make solving equations such as the Navier Stokes, with iterative solvers, a common day occurrence for even a mere laptop. However, many flow phenomena remain very hard to model accurately, turbulence being one of them [1]. Many relatively successful models exist today for modelling turbulence using Reynolds Averaged Navier Stokes (RANS) and Large Eddy Simulation (LES),which is why many of these techniques are used to significantly speed up the design process in industrial settings from aeroplanes to automobiles. Combustion and flame propagation is another area where CFD is expected to be the cornerstone for future development. With a clear shift towards sustainability, Hydrogen combustion is at the forefront of energy generation and propulsion due to its advantage of not producing CO2 [2]. Unfortunately, despite H2 being the simplest molecule, its combustion process is anything but that. It is strongly affected by thermodiffusive instabilities that produce a range of non-linearities in flame front propagation when coupled with turbulence [3],[4]. These non-linearities are critical to understand so that the use of Hydrogen in Gas Turbine (GT) and Internal Combustion Engines (ICE) is performed as efficiently as possible to maximise its energy output as well as ensure their safe operation. This is necessary since the energy stored in a given volume of hydrogen is much less than that of fossil fuels, making efficiency, the designer's number one priority. The complex 3D phenomena of thermodiffusive instability, and its coupling with turbulence, produce an increased flame speed and flame wrinkling, heavily dependent on the variation of local reaction rates. This means that, in the turbulent regime, the flame is highly irregular which is due to the combined contributions of turbulence and hydrogen's thermodiffusive instabilities. This produces tongue-like structures which penetrate in the unburned gas area which do not exist in turbulent combustion of traditional fuels such as methane. This highly irregular flow is very hard to model successfully which leads to the aim of this project. It is therefore envisioned to create accurate models of these phenomena to assist industrial design processes. Direct Numerical Simulations (DNS) simulations are extremely computationally expensive and cannot deemed feasible anywhere other than an academic setting. Therefore, by using existing DNS data and performing further simulations, it is planned to produce low order models that can be used in LES or even RANS. Furthermore, in addition to classical modelling approaches which are based on explicit equations [5], Machine Learning (ML) could be used [6]. ML is a powerful tool that allows relationships to be drawn between parameters arising from a dataset. Therefore, it could be possible to create the aforementioned models by training an algorithm based on the DNS data. This wouldmake use of the existing hardware of High-Performance Computing (HPC) available to the university which could then be used to retest the produced models against the initial dataset or other existing LES and RANS models. Thus, aiming to produce updated or even new models for Hydrogen turbulent combustion by the end of the project would pave the way for applying them in more complex settings and simulations and validate their performance against experimental data, in making the first step of implementing Hydrogen combustion in industrial applications.