目的 预测燃机透平叶片热障涂层孔隙率，加速热障涂层的研发及工艺优化，解决传统实验方法效率低、成本高的问题，为重型燃气轮机热障涂层研发及工业实际生产中的具体工艺参数调控提供一定指导。方法 采用MATLAB图像二值化处理技术计算陶瓷层的孔隙率数据，训练机器学习模型，预测不同工艺参数下热障涂层陶瓷层的孔隙率，并通过实验验证测试涂层的硬度和孔隙率。结果 Gradient Boosting Regression模型能够实现对热障涂层孔隙率的准确预测，喷涂功率、送粉率和喷涂距离对孔隙率的影响较大。机器学习具有一定的外延性，模型的R值(Related Coefficient，R)由0.834 4提高到0.943 0，R²值(Square of Related Coefficient，R²)从0.696 2提高到0.889 2，而MAE的值(Mean Absolute Error，MAE)从1.344 0降低到1.039 4，RMSE值(Root Mean Squared Error，RMSE)由1.881 0减少到1.712 8。随孔隙率的降低，等离子喷涂8YSZ陶瓷涂层的硬度由3.98 GPa增加到5.54 GPa，弹性模量由62.36 GPa提高到84.30 GPa。该模型准确预测了不同工艺下的涂层孔隙率。结论 喷涂功率、送粉率和喷涂距离决定了热障涂层的孔隙率，热障涂层的孔隙率与其硬度和弹性模量息息相关。本工作利用机器学习准确预测了不同工艺下的涂层孔隙率，证明机器学习算法在重型燃气轮机透平叶片热障涂层研发、工艺优化及生产中具有一定的应用前景。

Thermal barrier coating has been widely used in gas turbines, aircraft engines and other advanced power equipment due to its excellent performance in high temperature oxidation resistance, corrosion resistance and surface temperature reducing. The ceramic layer contains defects such as pores, microcracks, unmelted particles, microcracks, and interlayer interfaces. After long-term service at high temperature, the microstructure of the coating changes, which directly affects the insulation ability and mechanical properties of the coating, i.e., the porosity of the thermal barrier coating is directly related to its properties. However, the traditional experimental method has low research efficiency, there are many factors influencing the porosity of thermal barrier coatings, and the preparation process of thermal barrier coatings is very complex, with dozens of factors determining their performance. In order to further shorten the development and process optimization time of thermal barrier coatings, mathematical methods need to be used to establish models and analyze data to accelerate the development process. In this context, the branch of artificial intelligence data mining technology based on machine learning algorithms has gradually been introduced into the research and development of materials. Different from traditional mathematical fitting, data mining methods can establish nonlinear models and support simultaneous consideration of hundreds or even thousands of variables, allowing for extrapolation predictions without overfitting. In this work, the data mining technology was used to analyze the spraying data of thermal barrier coatings on heavy-duty gas turbine blades, establish multiple machine learning models that described the quantitative relationship between thermal barrier coating porosity and process parameters, compare the predictive effects of each machine learning, and test the predictive effects of the model using data from actual research and development processes. The MATLAB image binarization processing technology was used to calculate the porosity data of the ceramic layer, train a machine learning model to predict the porosity of the ceramic layer of the thermal barrier coating under different process parameters, and verify and test the hardness and porosity of the coating through experiments. According to the "No Free Lunch" theorem, no algorithm could be universally applicable to all situations. After trying with multiple machine learning algorithms, it was found that the Gradient Boosting Regression model was able to accurately predict the porosity of thermal barrier coatings, with spray power, powder feeding rate, and spray distance having the greatest impact on porosity. Machine science had a certain degree of extensibility, with the R value of the model increasing from 0.834 4 to 0.943 0, the R2 value increasing from 0.696 2 to 0.889 2, the MAE value decreasing from 1.344 0 to 1.039 4, and the RMSE value decreasing from 1.881 0 to 1.712 8. As the porosity decreased, the hardness value of plasma sprayed 8YSZ ceramic coating increased from 3.98 GPa to 5.54 GPa, and the Young‘s modulus increased from 62.36 GPa increased to 84.30 GPa. This model accurately predicts the porosity of coatings under different processes. Machine learning is used to predict the coating porosity under different processes accurately, it is proved that machine learning algorithms have certain application prospects in the research and development, process optimization, and production of thermal barrier coatings for heavy-duty gas turbine blades.