【摘要】 以全覆盖气膜冷却结构为研究对象，基于机器学习方法实现了温度场和热应力场的精细化和快速化预测。遴选孔倾斜角、孔复合角以及温比、吹风比为机器学习模型输入，基于拉丁超立方抽样和CFD仿真方法建立机器学习所需的学习样本；针对预测结构划分独立的机器学习网格节点(其网格数量远小于CFD网格数量)，并在每个节点上独立建立多层感知器神经网络模型直接预测该节点的温度/热应力；针对模型输入和输出参数之间的数值关联进行分析。研究结果表明：该模型用于预测固体域温度的均方差为41.21 K~2，平均绝对百分比误差为0.86%；预测固体域热应力的均方差为38.51 MPa~2，平均绝对百分比误差为0.15%。壁温随着复合角、吹风比、温比的增加而降低，随着孔倾斜角的增加而增大；热应力与壁温的变化趋势一致。所提出的数据压缩和预测策略针对每个网格节点独立建模，避免将节点坐标作为模型输入，提高了建模效率。更多还原

【Abstract】 The full coverage gas film cooling structure was taken as the research object,and the refined and rapid prediction of temperature field and thermal stress field was realized based on machine learning method.The hole inclination angle,hole compound angle,temperature ratio and blowing ratio were selected as the input of the machine learning model,and the learning samples required by machine learning were established based on Latin hypercube sampling and CFD simulation method.For the prediction structure,independent machine learning grid nodes were divided(the number of grids was far less than that of CFD grid),and a multi-layer perceptron neural network model was established independently on each node to directly predict the temperature/thermal stress of the node.The numerical correlation between the model input and output parameters was analyzed.The results show that the mean square error of the model for predicting the temperature of the solid domain is 41.21 K~2,and the mean absolute percentage error is 0.86%.The mean square error of the model for predicting the thermal stress of the solid domain is 38.51 MPa~2,and the mean absolute percentage error is 0.15%.The wall temperature decreases with the increase of the compound angle,the blowing ratio and the temperature ratio,and increases with the increase of the hole inclination angle.The thermal stress and the wall temperature have the same trend.The proposed data compression and prediction strategy is designed for each grid node independently,avoiding the node coordinates as the model input,and improving the modeling efficiency.更多还原