A neural network driven strategy for PyC interphase design in ceramic matrix composites

Ceramic matrix composites (CMCs), such as SiC f /SiC, mitigate the inherent brittleness of monolithic ceramics through fiber reinforcement, making them attractive for applications in extreme environments. Optimizing the fiber–matrix interphase is essential for improving fracture resistance but traditionally relies on costly trial-and-error. In this study, Phase-Field Cohesive Zone Modeling (PF-CZM) is employed to simulate the transverse fracture behavior of SiC f /PyC/SiC composites and generate physics-based training data. Based on these data, a dual neural-network framework consisting of a Forward Net and an Inverse Net is developed for PyC interphase optimization. The Forward Net serves as an efficient surrogate for the PF-CZM simulations, whereas the Inverse Net directly maps constituent properties and fiber geometry to the optimal PyC interphase thickness. Once trained, the Inverse Net directly predicts the optimal thickness without repeated iterative PF-CZM simulations during design optimization, thereby substantially reducing the computational cost. Closed-loop validation within the manufacturing-feasible domain (150–550 nm) demonstrates that the predicted thicknesses locate high-performance regions of the fracture-energy landscape. Ultimately, a dynamic design nomogram is established to provide fabrication guidelines. This interpretable, data-driven framework enables efficient microstructural customization for next-generation damage-tolerant CMCs.