关键词: 防热多孔材料, 多源异构数据融合, 渗透率, 深度学习, 多尺度表征

Abstract: Ablative thermal protection is an important thermal protection method for hypersonic vehicles. Porous pyrolytic car-bon materials are a type of ablative thermal protection material, whose permeability significantly influences transport characteristics. To address the issue of difficulty in obtaining empirical coefficients in formulas of permea-bility for thermal protection porous materials, a multi-source heterogeneous dataset was constructed, incorporating material microstructure images and macroscopic structural characteristic parameters (maximum flow and fractal dimension). A particle-based method named direct simulation of Monte Carlo was employed to calculate the per-meability of the microstructure. Based on this, three permeability prediction methods utilizing multi-source hetero-geneous data fusion strategies were proposed: a decision-level fusion strategy, a fusion strategy based on direct concatenation of multi-source data features, and a fusion strategy based on a cross-modal attention mechanism for multi-source data. By comparing the predictive performance of the three different fusion strategies, it was found that the feature-level fusion strategy based on the cross-modal attention mechanism, which captures the relation-ship between image convolutional features and structural characteristic parameters and dynamically adjusts the weights between structural images and structural characteristic parameters, achieved the best predictive perfor-mance. On the test set, the coefficient of determination (R2) was 0.95, and the mean absolute percentage error (MAPE) was 5.29%. Compared with single-source data-driven permeability prediction models, R2 improved by 6%, and MAPE decreased by 41%. This method enables efficient and accurate permeability prediction, providing tech-nical support for the refined design of thermal protection structures in practical hypersonic vehicles.

Key words: Ablative material, Multi-source heterogeneous data fusion, Permeability, Deep learning, Multi-scale characterization