Data-driven inverse design of C/C honeycombs via multiscale damage modeling

Carbon/carbon honeycomb sandwich structures (C/C-HSS) represent a next-generation solution for load-bearing platforms in space systems, offering superior mechanical performance, lightweight characteristics and excellent thermal stability. However, their design optimization poses a significant challenge due to highly nonlinear, multiscale interactions and conflicting objectives such as strength, stiffness, low weight, and thermal reliability. This study proposes a novel data-driven inverse design framework that integrates three key components: a multiscale thermo-mechanical-damage model, an interpretable deep learning surrogate, and multi-objective evolutionary optimization. A multiscale thermo-mechanical-damage model is first employed to accurately simulate the thermo-mechanical responses of C/C-HSS under varying thermal conditions, thus generating a comprehensive dataset. A deep neural network (DNN) surrogate is then trained on this dataset to predict eight critical structural properties with high accuracy (R² > 0.99, RMSE < 5 %). Feature importance is interpreted using SHapley Additive exPlanations (SHAP), revealing dominant design factors such as wall thickness and side length. Subsequently, a Non-dominated Sorting Genetic Algorithm II (NSGA-II) is employed to efficiently explore the design space and identify Pareto-optimal configurations that achieve significant weight reduction and high load-bearing capacity, while maintaining excellent thermal stability across a wide temperature range. This integrated approach achieves a 95 % reduction in computational expenditure compared to conventional finite element (FE) based optimization, while establishing a generalizable paradigm for multiscale inverse design of C/C-HSS operating under extreme aerospace thermal-mechanical conditions. The optimized configurations demonstrate 22 % mass reduction relative to baseline designs, while achieving uncompromised thermal stability, with computational efficiency exceeding traditional FE methods by an order of magnitude.