超声速燃烧与推进技术是实现高超声速飞行的关键问题之一。随着飞行马赫数的提高，CFD (Computational Fluid Dynamics)方法有效的补充了实验测量手段的不足。然而，随着湍流模型和数值格式的不断完善，超燃研究面临一些传统研究手段无法攻破的难题，且计算效率已经成为限制CFD方法用于工程问题的主要因素之一。另外，机器学习具有优异的大数据处理能力，其在建模、降维及优化设计等方面已经展现出巨大的潜力。在这样的研究背景下，本文以数值模拟为基础和主要数据来源，以ANN (Artificial Neural Network)模型和POD (Proper Orthogonal Decomposition)技术为主要研究手段，从微观到宏观，针对超临界条件下真实气体物性求解问题、全流场重构问题以及燃烧热释放速率预测问题开展相关研究。

    首先，在已有astroFoam计算平台的基础上，基于误差反向传播算法，本文成功搭建了一套具有三个隐含层结构的人工神经网络训练模型。另外，为了实现该ANN训练模型在CFD算例中的应用，本文成功地将上述发展的ANN训练模型框架耦合进了OpenFOAM计算平台，实现了模型训练过程和训练结果调用过程的灵活切换。

    针对超临界条件下，复杂碳氢燃料真实气体物性求解效率不高且计算代价过于昂贵的难题，本文提出一种高效而准确的ANN物性预测方法。并且以煤油燃料为例，对其3，5，10三种组分替代模型分别进行了ANN物性建模。结果分析显示，替代模型的组分数越多，对应ANN物性预测模型的网络拓扑结构越为复杂；每个ANN物性预测模型的预测精度均能满足要求，相关系数高达0.99以上；与ECS (Extend Corresponding State)方法相比，单独求解物性时，该ANN物性预测模型的效率可达到ECS求解效率的10⁴倍，且煤油替代模型的组分数越多，加速比越高。另外，本文通过对超临界煤油射流进行三维数值模拟，实现了该ANN物性预测模型与CFD算例的耦合。结果分析表明，该ANN物性计算模型可以准确地预测射流核心区长度，涡破碎位置，涡量分布以及温度和密度分布等，证实了该ANN物性预测模型用于三维数值模拟的准确可靠性；通过统计分析，发现对于本文CFD算例，采用ANN物性预测模型与ECS法则相比，计算加速比可达35.52％，且计算加速比随着计算网格总数量的增加而增大。

    高保真数值模拟用于优化设计不仅花费巨大、设计周期过长，而且不能做到实时预报。针对上述难题，本研究耦合计算流体力学、数据降维模型、统计学以及人工神经网络模型，提出一种基于ANN模型与CPOD (Common Proper Orthogonal Decomposition) 技术的准确且高效的全流场实时重构模型，用以预测包含丰富时空物理机制的全流场数据集，以提高在全设计空间范围内进行有效优化设计的效率。该模型实现过程主要包括：CPOD数据降维过程、基模态筛选过程、ANN训练过程以及预测重构性能分析过程等。并且以圆柱绕流和前台阶流动对该流场重构模型的可行性、准确性和高效性进行了验证。结果分析表明，整体上看该ANN全流场重构模型具有较高的准确性，但对于流场结构变化较为明显的设计参数范围，需要通过采样细化以加强对该区域的数据描述；该ANN全流场重构模型加速效果显著，在本算例中分别可达30和40倍，且随着问题复杂性的提高，采用该ANN全流场重构模型的计算加速效率将大大提高。

    优化燃烧释热分布是控制超燃、亚燃模态，提高燃烧效率和增加发动力推力的关键。但由于燃烧环境恶劣且燃烧过程包含太多复杂的物理化学过程，目前尚没有实验手段可以直接测量释热分布，且现存的预测模型也止步于定性分析。因此，基于POD数据降维技术和ANN数据驱动模型，本研究新提出一种燃烧热释放速率重构模型，以达到从实验可测化学发光元素方便、高效、定量的预测燃烧热释放速率空间分布的目的。主要实施过程包括：采用POD技术提取燃烧场的主要物理信息，随后通过一定的模态筛选方法达到对具有时空高分辨率的燃烧场数据进行降维、压缩的目的；通过ANN训练，建立降阶后HRR (Heat release rate)模态系数与化学发光元素OH模态系数之间定量、直接的映射关系；当给定新的化学发光元素OH的分布情况时，仅需要求解新状态下化学发光元素的模态系数，直接调用训练好的ANN模型预测出对应状态下燃烧释热率的模态系数，最终由新的HRR系数与原来的HRR基模态组合，即可重构新状态下HRR的分布情况。本研究以超声速氢氧燃烧扩散火焰为例，对该ANN-POD燃烧热释放速率重构模型的可靠性进行了验证。

     Supersonic combustion and propulsion technology is one of the key problems for achieving hypersonic flight. With the increasing of flight Mach number, CFD (Computational Fluid Dynamics) method can effectively supplement the insufficiency of experimental measurement methods. However, with the improvement of turbulence models and numerical algorithms, the investigation of supersonic combustion is encounted with chanllenges that can’t be solved by traditional research methods, and the computational efficiency has become one of the main factors limiting the application of CFD method to engineering problems. In addition, machine learning has excellent big data processing capabilities and has shown great potential in modeling, dimensionality reduction and optimization design. Under such a research background, this paper, based on numerical simulation and the main data source, takes ANN (Artificial Neural Network) model and POD (Proper Orthogonal Decomposition) technology as the main research means, from micro to macro, carries out related studies on the problem of solving the real gas physical properties under supercritical conditions, the reconstruction of the whole flow field and the prediction of combustion heat release rate.

     First, on the basis of the existing astrofoam computing platform and error back-propagation algorithm, this paper successfully builds a set of artificial neural network training model with three hidden layers. In addition, in order to realize the application of the ANN training model in the CFD example, this paper successfully coupled the ANN training model framework into OpenFOAM computing platform, and realized the flexible switch between the model training process and the loading process.

    Aiming at the problem that the calculation efficiency of real gas properties for complex hydrocarbon fuels is not high and the calculation cost is too expensive under supercritical condition, this paper proposes an efficient and accurate ANN method for predicting the physical properties of complex hydrocarbon fuels. Taking fuel of kerosene as an example, ANN physical property modeling of 3, 5, and 10 components surrogate model of kerosene fuel was carried out respectively. The results show that the more components of the surrogate model are, the more complex for the topological structure of the corresponding ANN model is. The prediction accuracy of each ANN physical property prediction model can meet the requirements, and the correlation coefficient is above 0.99. Compared with ECS (Extend Corresponding State) method, the efficiency of the ANN model can reach 10⁴ times faster than that of ECS method when solving the physical properties alone, and the more components of kerosene surrogate model are, the higher the acceleration ratio will be. In addition, through the three-dimensional numerical simulation of the supercritical kerosene jet, the coupling between the ANN physical property prediction model and the CFD example is realized. The results show that the ANN model for physical property prediction can accurately predict the lifting height, vortex breakage location, vorticity distribution, temperature and density distribution, etc., which proves the accuracy and reliability of the ANN model. Through statistical analysis, it is found that for the CFD case in this paper, the computational acceleration ratio can reach 35.52% by using this ANN model compared with ECS principle, and the computational acceleration ratio increases with the increase of grid size.

    High fidelity numerical simulation used to optimize design is not only costly, the design cycle is also time-consuming, and it can't realize real-time forecast. Thus, this study coupling computational fluid dynamics model, data dimension reduction, statistics and artificial neural network model, proposes an accurate and efficient full flow field real-time reconstruction model. This model is used to predict the instantaneous evolution process of the whole flow field containing rich spatiotemporal physical mechanisms within a large range of design space. The realization process of the model mainly includes: the dimensionality reduction process by applying the common POD technology, the basic mode modal screening process, the ANN training process and the prediction and reconstruction performance analysis process, etc. The feasibility, accuracy and efficiency of the flow field reconstruction model were verified by the cylinder flow and front step flow. The results show that the ANN flow field reconstruction model has a high accuracy generally; but for the region with obvious change of flow field structure, it is necessary to strengthen the data description of this region by sampling refinement. The acceleration effect of the ANN full-flow reconstruction model is significant, up to 30 and 40 times in this example, and with the increase of the complexity of the problem, the computational acceleration efficiency of the ANN full-flow reconstruction model will be greatly improved.

      Optimizing the distribution of HRR (Heat Release Rate) is the key to improve the performance of various combustors. However, limited by current diagnostic techniques, the spatial measurement of HRR in many realistic combustion devices is often difficult or impossible. HRR prediction is theoretically possible through establishing correlations between HRR and other quantities (e.g., chemiluminescence intensity) that can be experimentally determined; however, up to now, few universal correlations have been established. A novel artificial neural network approach was adopted to build the mapping relationship between the combustion heat release rate and the measurable chemiluminescent species. Proper orthogonal decomposition technology is used to extract the combustion physics and reduce the data of the spatial-temporally high-resolution combustion field. The correlation between the reduced-order HRR and chemiluminescent species is built using an ANN model. A unique segmentation approach was proposed to improve the training efficiency and accuracy. Validation in a supersonic hydrogen-oxygen nonpremixed flame proves the accuracy and efficiency of the proposed HRR reconstruction model based on the reduced-order POD method and data-driven ANN model.