Optimizing Material Property Prediction Precision via Global-Local Dual-Path Feature Extraction in a Deep Learning Framework

• Published in: 2024 4th Asia-Pacific Conference on Communications Technology and Computer Science (ACCTCS) pp. 162 - 166
• Main Authors: Chen, Qiaochuan, Han, Sifan, Han, Yuexing
• Format: Conference Proceeding
• Language: English
• Published: IEEE 24.02.2024
• Subjects: Coatings; Computer vision; Conductivity; Deep learning; Feature extraction; image analysis; thermal barrier coating; Thermal conductivity; Thermal expansion; Visualization
• DOI: 10.1109/ACCTCS61748.2024.00036

Thermal conductivity is a critical performance parameter for thermal barrier coating materials and a key thermal property. Shortening the cycle for predicting the thermal conductivity holds practical significance. Convolutional Neural Networks (CNNs) serve as the basis for many current studies aimed at processing images to predict material properties. However, in general, CNNs are limited in their receptive field due to their fixed-size convolution kernels, making it difficult to simultaneously extract both local and global information from images. In this paper, we develop a novel Global-Local Dual-Path Feature Extraction network termed GloLo-net, capable of extracting both global and local features from images and accurately predicting thermal conductivity based on material image data. A newly designed multiscale attention fusion module can comprehensively capture visual features of the material's microstructure. Experiment results show the deep learning neural network outperforms many existing data-driven methods.