Abstract: To solve the problems of lack of mechanical explanatory, weak generalization ability and limited engineering application due to the ‘black box’ characteristics of traditional data-driven models, this paper proposes a Structured Transfer Learning-Knowledge Inference Neural Network (STL-KINN) method to use a collaborative modeling strategy of mechanism embedding and knowledge transfer. Improve the explainability of black-box models such as neural networks and their generalization processes. Firstly, based on the Hashin criterion, typical failure modes such as fiber fracture and matrix cracking are directly embedded into the network architecture, and structured polynomial functions are designed, and each order term has a clear physical meaning, so that the model weights and physical laws can be mutually verified. Secondly, in view of the defect of non-quantifiable explanation of the hyperparameter transfer process, the weight threshold in the structured polynomial function is fine-tuned in combination with the difference between the source domain and the target domain, and the changes of the hyperparameters can be quantified visually. Finally, the prediction accuracy of the compressive strength of X850 material laminate is reduced by 85.882% under the condition of sufficient sample size, and the MRE is reduced by 59.062% after transfer learning, which provides a new idea for improving the generalization ability and engineering applicability of the enhanced model.