陶瓷涂层具有耐高温、耐腐蚀、耐磨损等特性, 其热膨胀系数和热导率等参数与其性能息息相关. 为解决陶瓷涂层性能实验成本高、测试困难等问题, 提出了陶瓷涂层材料多模态数据表征学习的性能预测方法. 首先利用高斯混合模型虚拟样本生成(Gaussian mixture model virtual sample generation, GMMVSG)算法生成符合真实陶瓷涂层数据分布的样本来扩充数据集; 其次利用卷积神经网络 VGG16 对陶瓷涂层的显微结构图像数据进行特征提取, 利用 TabNet 对结构化数据进行特征提取, 将提取到的图像数据特征与结构化数据特征融合; 最终根据多模态数据表征建立基于K-最近邻(K-nearest neighbor, KNN)、支持向量机回归(support vector regression, SVR)和多层感知机(multi-layer perceptron, MLP) 3 种机器学习算法的预测模型, 对陶瓷涂层的性能指标, 即热膨胀系数和热导率进行了预测. 实验结果表明: 提出的多模态数据表征学习模型的预测结果要优于单模态数据表征学习模型, 其中基于 MLP 算法训练的多模态数据表征学习模型对陶瓷涂层性能的预测效果最好; 在测试集中, 对陶瓷涂层热膨胀系数预测的平均绝对误差(mean absolute error, MAE)和均方误差(mean square error, MSE)分别为 0.026 6 和 0.001 7, 对热导率预测的 MAE 和 MSE 分别为 0.017 9 和 0.000 7. 所提出的陶瓷涂层材料多模态数据表征学习方法有效融合了结构化数据与非结构化数据, 联合学习了各模态数据的潜在共享信息, 成功提升了对陶瓷涂料层材料性能预测的准确度.

Ceramic coatings have excellent temperature resistance, corrosion resistance, and wear resistance, among other advantages. Their thermal expansion coefficient and thermal conductivity are two properties directly related to their performance. To address the issues of high experimental costs and challenging test conditions, we propose a method to predict the performance of ceramic coating materials based on multimodal data representation learning. To enlarge the data set, this method uses the Gaussian mixture model virtual sample generation (GMMVSG) algorithm to generate samples that match the real ceramic-coating data distribution. The method extracts micro-structural image data's features using the very deep convolutional neural network VGG16, extracts structured data's features using TabNet, and fuses the features of the extracted image data with those of the structured data. the final prediction models based on three machine learning algorithms-K-nearest neighbor (KNN), support-vector-machine regression (SVR), and multi-layer perceptron（MLP）—are established by using multimodal data representation to predict the thermal expansion coefficient and thermal conductivity of the performance index of ceramic coatings. The experimental results show that the proposed multimodal-data representation-learning model has a better prediction performance than that of the single-modal-data machine-learning model, and that the former model based on the MLP can most accurately predict ceramic coating performance. In the test set, the mean absolute and mean square errors for the prediction of the thermal expansion coefficient are 0.026 6 and 0.001 7, respectively, and the mean absolute and mean square errors for the prediction of thermal conductivity are 0.017 9 and 0.000 7, respectively. Our proposed learning method for multimodal data representation of ceramic coating materials effectively combines structured and unstructured data to learn both types of modal data with potentially shared information and successfully improves the pred.