Machine learning-based multi-objective optimization and physical-geometrical competitive mechanisms for 3D woven thermal protection composites

3D woven composite materials are prime candidates for thermal protection due to their significant thermophysical properties, which necessitates accurate prediction of these properties and precise meso-structural design. This study introduces a fusion framework that integrates machine learning-based sensitivity analysis and multi-objective optimization design. Features selected for model training included meso-scale geometric structures and the physical properties of the constituent components. Through extensive numerical calculations, three datasets comprising a total of 5650 data points were established to train machine learning models for predicting three essential equivalent performances: thermal conductivity, density, and surface emissivity. The established machine learning models were utilized for sensitivity analysis and multi-objective optimization design. Among the regression models evaluated (Lasso, RF, SVR, and MLP), the MLP model demonstrated superior performance for predicting equivalent thermal conductivity, while the SVR model excelled in predicting equivalent density and surface emissivity. Sobol sensitivity analysis revealed that the meso-structural parameters had total-order sensitivity indices of 65.0 % for equivalent thermal conductivity, 71.2 % for equivalent density, and 21.2 % for equivalent surface emissivity. The NSGA-II multi-objective optimization identified a phenomenon of fakePareto fronts, by reducing the design space by 5 %, this issue was mitigated. Considering dataset generation and optimization, the fusion framework requires only 22.4 % of the cost compared to traditional numerical methods. Meanwhile, in the process of machine learning accelerated-optimization, boundaries should be carefully selected.