Emissivity Prediction of Multilayer Film Radiators by Machine Learning using an Ultrasmall Dataset

Tianzhu Fan

Ying LiEmail

J. Mike Walker ’66 Department of Mechanical Engineering, Texas A&M University, College Station, TX, 77843, USA

Abstract

Multilayer films consisting of alternating layers of materials are a popular structural design to achieve efficient radiative cooling. To the best of our knowledge, this is the first study to employ machine learning to forecast the emissivity of multilayer film radiators using an ultra-small dataset (32 cases) from published articles. The physical parameters of each layer and the sequence of the coatings were extracted as features of the prediction models. Machine-learning algorithms, including gradient boosting and artificial neural networks, have been applied to build prediction models. Subsequently, the K-fold cross-validation method was employed to evaluate the performance of the models. As a result, the optimized algorithms overcame the weakness of the ultra-small dataset and provided a binary classification accuracy of 87.5% to predict whether a multilayer film structure was an “efficient emitter,” whose hemispherical emissivity was higher than 0.85 in the atmospheric window. In the regression model, the maximum correlation coefficient between the predicted and experimental emissivities achieved 0.76, indicating a strong positive correlation. Consequently, the differences between the predicted and experimental values were less than 10%. Also, feature engineering was implemented to compare the importance of physical parameters on the predicted emissivities. 

Emissivity Prediction of Multilayer Film Radiators by Machine Learning using an Ultrasmall Dataset