A Dynamic‐Weighted Deep Transfer Learning Framework for Thermal Conductivity Prediction and Analysis

ABSTRACT To overcome the limitations of small‐sample data in establishing microstructure–property linkages, this study introduces a deep transfer learning framework with a dynamic weighting mechanism. By transferring the perceptual capabilities of the ResNet‐18 model and adaptively fusing them with material domain knowledge, the model effectively captures complex features such as phase distribution. Using silicon nitride ceramics as the primary research object, the framework achieves an average cross‐validation prediction accuracy ( R 2 ) of 0.73, representing a 104.3% relative improvement compared to the traditional CNN framework, and the optimal model reaches an accuracy of R 2 = 0.89. Furthermore, this framework also demonstrates exceptional predictive accuracy on silicon carbide ceramics ( R 2 = 0.84) and sintered nano silver ( R 2 = 0.93), indicating its strong generalization capabilities. By employing multilevel gradient‐weighted class activation mapping (Grad‐CAM) and sliding occlusion analysis, the decision‐making process of the model is elucidated, thereby validating the logical soundness of its predictions. Additionally, symbolic regression is utilized to identify the influence of different microstructural features on thermal conductivity and to establish their quantitative relationships. This research holds broad application prospects in the rapid development and design of thermal management materials, analysis of material microstructure images, and the establishment of structure–performance relationships between microstructural features and macroscopic properties.