IMT Physics-based and Data-driven Modelling of pollutant Emissions from Engines

我感兴趣的项目是广告和它的标题为“基于物理和数据驱动的建模污染物排放的发动机”。该项目涉及模拟燃气涡轮机发动机的碳烟颗粒排放。碳烟是燃气涡轮机发动机产生的主要污染物之一，因此碳烟的建模和预测对于下一代低排放燃气涡轮机和内燃机的发展至关重要。为了研究燃气涡轮机发动机中的碳烟颗粒演变，需要四个不同的部件：背景湍流模型，气相燃烧模型，通过各种微观过程如初始影响烟灰颗粒的物理化学机制模型，生长和氧化以及颗粒演化动力学模型。模拟碳烟排放的最准确方法是通过直接数值模拟（DNS）其直接求解非定常的Navier-Stokes方程，并且能够解决湍流中烟灰颗粒的小尺度相互作用，但是这些解决方案伴随着大量的计算费用。由于这个原因，其他相对较少的计算昂贵的模型已被广泛使用，如大涡模拟（LES）。尽管大涡模拟被广泛用于模拟湍流反应流，但要实现烟尘颗粒、化学物质和湍流之间的小尺度相互作用的精确模拟仍然是一个艰巨的挑战。因此，这个博士项目的目的是解决三个问题，在大涡模拟时遇到的烟尘的形成和演变，以开发一个增强的大涡模拟模型，以准确地预测模型燃气涡轮机燃烧室的烟尘排放。本文主要研究了以下三个问题：1.在非结构网格上发展一种一致的LES/概率密度函数（PDF）方法，以精确地表征湍流之间的小尺度相互作用，通过求解描述火焰结构和气相前驱体演化的标量的联合子滤波PDF方程以及数密度函数（NDF）的矩，2.将单个物质的分子扩散系数并入PDF求解器中以研究分辨的微分扩散对成核的影响，3.评估碳烟特性对碳烟前体化学的敏感性以及对用于重建碳烟颗粒NDF的矩量法的选择的敏感性。新的增强LES/PDF-本文将采用CFD模型对德国DLR研制的模型燃气涡轮机燃烧室进行数值模拟。利用DLR提供的数据集对结果进行了验证，该数据集是在高压燃气涡轮机燃烧室中使用高速激光诊断实验产生的。将在湍流壁射流扩散火焰上运行DNS，获得的有价值的数据集将用于训练基于卷积神经网络（CNN）的降阶模型，用于预测燃气涡轮机发动机的碳烟排放。目的是将基于物理的模型（通过实现先前目标获得）和CNN模型联合收割机结合起来，开发CNN辅助的混合基于物理的模型，该模型能够以更低的计算成本准确预测碳烟排放。

The project that I'm interested is advertised and its titled "Physics-based and Data-driven Modelling of pollutant Emissions from Engines ". The project involves in modeling soot particle emissions from gas turbine engines. Soot is a major pollutant produced by gas turbine engines therefore the ability to model and predict soot is crucial to the development of next generation low emission gas turbine and internal combustion (IC) engines.Modeling soot emissions a particularly challenging problem due to its small scale interactions between turbulence, particle dynamics and chemistry. To study soot particle evolution in gas turbine engines, it requires four different components: model for background turbulent flow, model for gas phase combustion, model for physico-chemical mechanisms that effects the soot particles by various micro-process like inception, growth and oxidation and model for particle evolution dynamics.The most accurate way to simulate soot emissions is through direct numerical simulations (DNS) which directly solves the unsteady Navier-Stokes equations and is capable of resolving small scale interactions of soot particles in turbulent flows but these solutions come with a great deal of computational expense. Due to this reason other relatively less computationally expensive models have been extensively used, such as the large eddy simulations (LES). Even though LES is widely employed to model turbulent reacting flows, it still remains a formidable challenge to achieve accurate modeling of small scale interactions between soot particles, chemistry and turbulence. Therefore this PhD project aims to address three issues encountered in LES when modeling soot formation and evolution in order to develop an enhanced LES model to accurately predict soot emissions in a model gas turbine combustor. The three main issues addressed are listed below.1.Develop a consistent LES/probability density function (PDF) approach on unstructured meshes to accurately characterize small scale interactions between turbulence, soot and chemistry in a gas turbine model combustor by solving the joint sub-filter PDF equation of the scalars used to describe the flame structure and gas-phase precursor evolution as well as the moments of number density function (NDF) of soot particles2.Incorporate molecular diffusivities of individual species into the PDF solver to study the effects of resolved differential diffusion on nucleation, growth and oxidation of soot particles.3.Assessing the sensitivity of soot characteristics to soot-precursor chemistry and to the choice of method of moments (MOM) that is used to reconstruct the NDF of soot particles.The new enhanced LES/PDF-MOM model will be used to simulate a model gas turbine combustor developed by DLR Germany. The results will be validated using a dataset provided by DLR, which was experimentally produced using high speed laser diagnostics in a high pressure gas turbine combustor.A DNSs will be run on turbulent wall jet-diffusion flame and the valuable dataset obtained will be used to train a convolutional neural network (CNN) based reduced order model for predict soot emissions from gas turbine engines. The aim is to combine the physics-based model (obtained from achieving the previous objective) and the CNN model to develop a CNN assisted hybrid physics-based model that is capable of accurately predicting soot emission at a reduced computational cost.