以超临界流体作为循环工质的能源利用系统广泛应用于动力工程、航天技术、热泵和制冷等领域，超临界压力流体对流换热逐渐成为传热学领域的前沿研究热点。然而，在准临界温度附近，超临界流体剧烈的变物性会引起浮升力、流动加速或其二者的耦合作用，导致换热恶化或强化等异常现象，给换热系统的设计和性能预测带来了很大挑战，阻碍了相关技术的发展和应用。一方面，现有理论模型还不能准确描述超临界压力流体的湍流换热性能。另一方面，能够相对准确地预测换热性能的实验或直接数值模拟（DNS）方法往往成本较高、研究周期较长，无法满足日常研究和工业应用的即时需要；能够快速预测换热性能的经验关联式或相对计算量较小的数值模拟方法适用的工况范围有限，预测精度一般为75%甚至更低。上述方法的缺陷使得换热系统的设计以及对换热性能的预测变得更加困难，因此，如何获得准确且快速的超临界压力流体对流换热性能预测方法是目前尚待解决的重要研究问题。 本文整理汇总了本课题组超临界压力流体湍流换热的实验和DNS结果，包含CO2和正癸烷两种工质共计超过90000个数据点。通过对比模型拟合能力和预测精度选择了机器学习中的LightGBM算法用于构造预测超临界压力流体换热性能的替代模型，其中包括：适用于单工质、单因素影响的模型，适用于单工质、多因素影响的模型，以及适用于多种工质、多种因素影响的统一模型。结果表明：替代模型的平均误差均小于5%，显著低于已有经验关联式的平均误差，其泛化性和鲁棒性均得到验证，误差波动不超过±1%，且适用于多种工质在多因素耦合影响下的复杂工况。通过对数据集中特征的扩充和筛选构造了只使用无量纲参数的替代模型，在不引入任何与壁面流体相关的参数时，模型预测误差约为11%，相比于已有经验关联式依赖壁面温度作为定性温度进行替代计算而言，在工业应用中具有更强的实用性。本文还对模型的内部结构进行分析，尝试解释机器学习“黑箱”模型中的计算规则，基于对特征重要性的分析和现有理论研究构造了新的无量纲数A和B，进一步将模型的预测误差降低了约0.3%。这一分析对综合利用已有研究数据、构造融合热物理先验知识的机器学习方法、提高模型可解释性以及借助所构造的模型来推动超临界流体传热研究发展具有一定的启发性，对攻克传热学中的复杂问题、构造准确的预测模型具有创新意义。

With its excellent heat and mass transfer characteristics, supercritical fluids have shown broad application prospects and are widely used in power engineering, aerospace technology, heat pump, refrigeration, and other fields. Energy utilization system using supercritical fluids as working medium has attracted great attention. However, near the pseudo-critical temperature, supercritical fluid has strong thermo-property variations that further causes buoyancy effect, flow acceleration effect and the coupling effect of the two effects, resulting in heat transfer deterioration and other anomalies. This makes the mechanism of turbulent heat transfer more complicated, brings great challenges to the related theoretical research and heat transfer system design, and hinders its development and applications. Existing theory is yet incapable of accurately describing the mechanism of supercritical pressure fluid turbulent heat transfer. Experimental methods and direct numerical simulation (DNS) method can accurately analyze and predict the local heat transfer performance with higher cost and longer research cycle. Empirical correlations and numerical simulation method, which are only applicable to limited range of working conditions and consistent to the experimental results only qualitatively in most cases, can predict the heat transfer performance rapidly but less accurately with lower calculational cost. The defects of the above methods make it more difficult to design the heat transfer system and predict the heat transfer performance. Methods that yield heat transfer performance prediction with high accuracy at low cost in both time and computations are thus needed for further applications. Consequently, how to acquire accurate and fast methods to predict the convective heat transfer performance of supercritical pressure fluids is an important research problem to be solved at present. This paper summarizes the experimental and DNS results of turbulent heat transfer of supercritical pressure fluids, including more than 90,000 data points and two working fluids, which are CO2 and decane. By comparing the model fitting ability and the accuracy of prediction of three different machine learning algorithms, LightGBM is selected to construct the supercritical pressure fluid surrogate models, including: models that are suitable for cases where only one working medium and one factor are involved; models that are suitable for cases where single medium and multiple factors are involved; and models that are applicable to multiple media and multiple factors. The results show that the average error of surrogate models is less than 5%, which is significantly lower than that of empirical correlations. The generalization ability and robustness of surrogate models are verified and the error fluctuates within a range of ±1%. Surrogate models thus built are suitable for the complicated working conditions where multiple media and multiple factors are involved. The dimensionless model using only dimensionless parameters is fitted by selecting and expanding the features of the dataset. When no parameters related to wall fluid are introduced, the prediction error of the model is about 11%, compared with the empirical correlations that relies greatly on the wall temperature to carry out the calculation, the prediction performance of the model is significantly improved and it's more practical in industrial applications. This paper also analyzes the internal structure of surrogate models, attempting to explain the calculation rules of the machine learning black box model. Two new dimensionless numbers A and B are proposed based on the analysis of feature importance and existing theoretical research. By introducing them to the features of the dataset, the prediction error of the model is further reduced by 0.3%. This paper may inspire future theoretical analysis using the existing research data by constructing machine learning models which integrate prior knowledge of thermophysics. The overall interpretability of machine learning models is also improved. It also has innovative significance for solving the complex problems in heat transfer.