Carbon fiber reinforced polymers (CFRPs) are widely used in several critical areas. Due to the large design parameter space, the performance prediction of CFRPs can effectively avoid the shortcomings of traditional design and optimization methods, which are time-consuming and high cost. However, the lack of standardized data makes the existing literature mining methods rarely applicable to the field of CFRPs. Therefore, we propose an entity identification method named SRGN, which extracts key information from literature using heterogeneous graphs, and establishes a CFRPs literature dataset CompMatLitDS, which contains 13 types of entities related to material composition, preparation process parameters, sample thickness, test methods, and properties. To validate the effectiveness of SRGN method, experiments on CompMatLitDS achieved a Precision of 95.83%. Further experiments on the public material dataset, MatScholar, demonstrated a modest improvement in accuracy compared to other commonly used methods in the material field, further demonstrating the validity of the SRGN method. Based on the entity information obtained by mining, along with data sorting and filtering, nine critical features affecting the flexural strength of CFRPs under curing-dominated processing conditions are identified. Four machine learning models are employed to predict this property. Both the random forest and XGBoost models achieved an R 2 of 0.93, while the random forest model demonstrated the highest overall performance by exhibiting a significantly lower mean absolute error. Feature analysis identified key factors and their corresponding thresholds under curing-dominated processing conditions, which can guide the optimization of material composition and processing techniques, thus advancing both the innovation and practical application of CFRPs.While the proposed SRGN method and CompMatLitDS dataset are designed for generalized CFRPs, the detailed performance prediction and feature analysis specifically focus on epoxy-based systems due to their high data density in current literature. The generalization of the model to other thermoplastic matrices or more complex processing conditions remains to be further validated as more diverse data becomes available in the future.
Zhang et al. (Mon,) studied this question.