超临界CO2布雷顿循环具有效率高、紧凑度大的优点，在能源动力领域具有广阔的应用前景。超临界压力流体在微通道换热器内的湍流换热变化规律复杂，严重影响超临界系统的安全性、稳定性和效率。然而，超临界压力流体湍流换热涉及到湍流输运与热量输运的高度耦合影响，受限空间和强变物性诱发复杂的局部换热恶化和换热强化现象，传统的换热预测方法存在误差大、泛化性差等问题。因此，深入研究微细通道内超临界压力流体湍流动力学特性，揭示微细通道内超临界压力流体湍流换热机理，发展高效准确的换热预测方法，对于超临界CO2布雷顿循环的发展具有重要的研究意义。本文从认识微细通道内超临界压力CO2湍流换热的微观机理出发，采取机理认识与机器学习相融合的学术思路构建新型湍流换热模型，开展了可视化实验、直接数值模拟与机器学习研究。全文主要结论和认识如下：开展了微细通道内超临界压力CO2流动换热的可视化实验研究，搭建了超临界压力流体Micro-PIV可视化实验系统，实现了对超临界压力CO2速度场和温度场同步测量。结果表明，加热条件下微槽道内超临界压力CO2沿流向平均流动加速，加速湍流发生衰减和再生成，导致局部换热恶化与恢复。通过超临界压力CO2湍流换热的直接数值模拟研究，揭示了变物性和受限空间约束下超临界压力流体湍流衰减与再生成的规律和机理，发现微米尺度效应抑制了近壁涡向中心区域的对流和抬升。揭示了浮升力和流动加速导致湍流动能向平均动能的逆输运机制，阐明了局部加速强化流向涡拉伸项进而破坏湍流涡的各向异性生成机制，提出了定量描述超临界压力流体湍流换热规律的无量纲数。利用机器学习挖掘了浮升力与流动加速影响下超临界压力流体流动换热雷诺平均过程的关键物理量，发现在强浮升力作用下已有雷诺比拟模式失效机理。基于深度神经网络方法改进了超临界压力流体湍流换热模型，改进模型相较于原始模型在定量上提高了计算精度，壁面温度预测误差小于10%，并且模型对于不同超临界流体具有很好的泛化性，收敛速度提升20%。以高超声速飞行器RBCC燃烧室第三流体主动冷却方案为对象，提出冷却方案一维分析方法与燃烧室-冷却通道三维耦合数值模拟方法。研究表明，回热循环可以改善流量分配不均匀的问题，提升CO2的热沉利用率，整体循环效率较高。

Supercritical CO2 Brayton cycle is an advanced power cycle with high efficiency and high compactness, which has a broad application prospect in the field of energy and power. The turbulent heat transfer of supercritical pressure fluid in microchannel heat exchanger is complicated, which seriously affects the safety, stability and efficiency of supercritical system. However, the turbulent heat transfer of supercritical pressure fluid involves the highly coupled influence of turbulent and heat transport, and the constrained space and strong physical properties induce complex local heat transfer deterioration and heat transfer enhancement. The traditional heat transfer prediction method has problems such as large error and poor generalization. Therefore, it is of great significance for the development of supercritical CO2 Brayton cycle to deeply study the turbulent dynamic characteristics of supercritical pressure fluids in microchannels, reveal the turbulent heat transfer mechanism of supercritical pressure fluids in microchannels, and develop efficient and accurate heat transfer prediction methods.Based on the understanding of the underlying mechanism of the turbulent heat transfer of supercritical CO2 in the microchannels, this paper adopts the academic idea of combining mechanism understanding and machine learning to build a new heat transfer model, and carries out the research of visualization experiment, direct numerical simulation and machine learning. The main conclusions and understandings are as follows:A visualization experimental study on the heat transfer mechanism of supercritical pressure CO2 flow in a micro-channel was carried out, and a Micro-PIV visualization experimental system was built for supercritical pressure fluid, realizing synchronous measurement of the velocity field and temperature field of supercritical pressure fluid. The results show that under heating conditions, the supercritical pressure fluid in the microchannel flows along the direction of the accelerated flow, and the accelerated turbulence occurs local attenuation and regeneration, leading to the deterioration and recovery of local heat transfer.Direct numerical simulation (DNS) was used to investigate turbulent heat transfer in supercritical pressure CO2. The law and mechanism of turbulent decay and regeneration were revealed under varying physical properties and constrained space. It was found that the micron scale effect inhibited the convection and uplift of near-wall vortex towards the center region. The inverse transport mechanism of turbulent kinetic energy to average kinetic energy caused by buoyancy and flow acceleration was revealed, the anisotropic generation mechanism of turbulent vortices was elucidated by local acceleration strengthening of flow vortex tension term, and the anisotropic generation mechanism of turbulent vortices was proposed.Machine learning was used to explore the key physical quantities of Reynolds mean process of heat transfer in supercritical pressure fluids under the influence of buoyancy and flow acceleration, and it was found that the Reynolds‘s analogy mode failured with the effect of strong buoyancy. The turbulent heat transfer model of supercritical pressure fluid was improved based on the deep neural network method. Compared with the original model, the calculation accuracy of the improved model was improved quantitatively. The prediction error of wall temperature was less than 10%, and the convergence rate of the improved model was improved by 20%.Taking the third fluid cooling thermal protection solution of hypersonic vehicle RBCC combuster as the object, the one-dimensional analysis method and three-dimensional coupling numerical simulation method of active cooling solution are proposed. The results showed that the heat sink utilization rate of CO2 can be improved by reheating cycle to improve the uneven flow distribution, and the overall cycle efficiency is higher.